patch-2.2.16 linux/drivers/net/acenic.c
Next file: linux/drivers/net/acenic.h
Previous file: linux/drivers/net/Space.c
Back to the patch index
Back to the overall index
- Lines: 3260
- Date:
Wed Jun 7 14:26:43 2000
- Orig file:
v2.2.15/linux/drivers/net/acenic.c
- Orig date:
Wed May 3 17:16:40 2000
diff -urN v2.2.15/linux/drivers/net/acenic.c linux/drivers/net/acenic.c
@@ -2,7 +2,7 @@
* acenic.c: Linux driver for the Alteon AceNIC Gigabit Ethernet card
* and other Tigon based cards.
*
- * Copyright 1998, 1999 by Jes Sorensen, <Jes.Sorensen@cern.ch>.
+ * Copyright 1998-2000 by Jes Sorensen, <Jes.Sorensen@cern.ch>.
*
* Thanks to Alteon and 3Com for providing hardware and documentation
* enabling me to write this driver.
@@ -17,14 +17,24 @@
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
- * Additional work by Pete Wyckoff <wyckoff@ca.sandia.gov> for initial
- * Alpha and trace dump support.
+ * Additional credits:
+ * Pete Wyckoff <wyckoff@ca.sandia.gov>: 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 <davem@redhat.com>: 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 <Matt_Domsch@dell.com>: Detect 1000baseT cards
*/
-#define PKT_COPY_THRESHOLD 300
-
+#include <linux/config.h>
#include <linux/module.h>
-
+#include <linux/version.h>
#include <linux/types.h>
#include <linux/errno.h>
#include <linux/ioport.h>
@@ -36,6 +46,13 @@
#include <linux/init.h>
#include <linux/delay.h>
#include <linux/mm.h>
+
+#undef ETHTOOL
+#undef INDEX_DEBUG
+
+#ifdef ETHTOOL
+#include <linux/ethtool.h>
+#endif
#include <net/sock.h>
#include <net/ip.h>
@@ -45,20 +62,19 @@
#include <asm/byteorder.h>
#include <asm/uaccess.h>
-#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"
+#ifdef CONFIG_ACENIC_OMIT_TIGON_I
+#define ACE_IS_TIGON_I(ap) 0
+#else
+#define ACE_IS_TIGON_I(ap) (ap->version == 1)
+#endif
#ifndef PCI_VENDOR_ID_ALTEON
#define PCI_VENDOR_ID_ALTEON 0x12ae
-#define PCI_DEVICE_ID_ALTEON_ACENIC 0x0001
+#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
@@ -71,18 +87,107 @@
* They used the DEC vendor ID by mistake
*/
#ifndef PCI_DEVICE_ID_FARALLON_PN9000SX
-#define PCI_DEVICE_ID_FARALLON_PN9000SX 0x1a
+#define PCI_DEVICE_ID_FARALLON_PN9000SX 0x1a
#endif
#ifndef PCI_VENDOR_ID_SGI
-#define PCI_VENDOR_ID_SGI 0x10a9
-#define PCI_DEVICE_ID_SGI_ACENIC 0x0009
+#define PCI_VENDOR_ID_SGI 0x10a9
+#endif
+#ifndef PCI_DEVICE_ID_SGI_ACENIC
+#define PCI_DEVICE_ID_SGI_ACENIC 0x0009
+#endif
+
+#ifndef wmb
+#define wmb() mb()
+#endif
+
+#ifndef __exit
+#define __exit
+#endif
+
+#ifndef SMP_CACHE_BYTES
+#define SMP_CACHE_BYTES L1_CACHE_BYTES
+#endif
+
+
+#if (LINUX_VERSION_CODE < 0x02030d)
+#define pci_resource_start(dev, bar) dev->base_address[bar]
+#elif (LINUX_VERSION_CODE < 0x02032c)
+#define pci_resource_start(dev, bar) dev->resource[bar].start
+#endif
+
+#if (LINUX_VERSION_CODE < 0x02030e)
+#define net_device device
+#endif
+
+#if (LINUX_VERSION_CODE >= 0x02031b)
+#define NEW_NETINIT
+#endif
+
+#if (LINUX_VERSION_CODE < 0x02032a)
+typedef u32 dma_addr_t;
+
+static inline void *pci_alloc_consistent(struct pci_dev *hwdev, size_t size,
+ dma_addr_t *dma_handle)
+{
+ void *virt_ptr;
+
+ virt_ptr = kmalloc(size, GFP_KERNEL);
+ *dma_handle = virt_to_bus(virt_ptr);
+ return virt_ptr;
+}
+#define pci_free_consistent(cookie, size, ptr, dma_ptr) kfree(ptr)
+#define pci_map_single(cookie, address, size, dir) virt_to_bus(address)
+#define pci_unmap_single(cookie, address, size, dir)
+#endif
+
+#if (LINUX_VERSION_CODE < 0x02032b)
+/*
+ * SoftNet
+ */
+#define dev_kfree_skb_irq(a) dev_kfree_skb(a)
+#define netif_wake_queue(dev) clear_bit(0, &dev->tbusy)
+#define netif_stop_queue(dev) set_bit(0, &dev->tbusy)
+
+static inline void netif_start_queue(struct net_device *dev)
+{
+ dev->tbusy = 0;
+ dev->interrupt = 0;
+ dev->start = 1;
+}
+
+#define ace_mark_net_bh(foo) mark_bh(foo)
+#define netif_queue_stopped(dev) dev->tbusy
+#define netif_running(dev) dev->start
+#define ace_if_down(dev) {do{dev->start = 0;}while (0);}
+#else
+#define NET_BH 0
+#define ace_mark_net_bh(foo) {do{} while(0);}
+#define ace_if_down(dev) {do{} while(0);}
#endif
+
+#define ACE_MAX_MOD_PARMS 8
+#define BOARD_IDX_STATIC 0
+#define BOARD_IDX_OVERFLOW -1
+
+
+#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"
+
/*
* This driver currently supports Tigon I and Tigon II based cards
- * including the Alteon AceNIC and the 3Com 3C985. The driver should
- * also work on the NetGear GA620, however I have not been able to
- * test that myself.
+ * 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
@@ -147,6 +252,12 @@
* 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
@@ -160,46 +271,146 @@
*
* 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 independant 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.
*/
/*
- * Default values for tuning parameters
+ * 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 DEF_TX_RATIO 31
-#define DEF_TX_COAL 1000
-#define DEF_TX_MAX_DESC 40
-#define DEF_RX_COAL 1000
-#define DEF_RX_MAX_DESC 20
-#define DEF_TRACE 0
-#define DEF_STAT 2 * TICKS_PER_SEC
-
-static int link[8] = {0, };
-static int trace[8] = {0, };
-static int tx_coal_tick[8] = {0, };
-static int rx_coal_tick[8] = {0, };
-static int max_tx_desc[8] = {0, };
-static int max_rx_desc[8] = {0, };
-static int tx_ratio[8] = {0, };
+#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
-static const char __initdata *version = "acenic.c: v0.33a 08/16/99 Jes Sorensen (Jes.Sorensen@cern.ch)\n";
-static struct device *root_dev = NULL;
+/*
+ * 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 + 2 + 16)
+#define ACE_STD_BUFSIZE (ACE_STD_MTU + ETH_HLEN + 2+4+16)
+#define ACE_JUMBO_BUFSIZE (ACE_JUMBO_MTU + ETH_HLEN + 2+4+16)
+
+/*
+ * 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 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
+
+#define TX_COAL_INTS_ONLY 0 /* seems not worth it */
+#define DEF_TRACE 0
+#define DEF_STAT (2 * TICKS_PER_SEC)
+
+static int link[ACE_MAX_MOD_PARMS] = {0, };
+static int trace[ACE_MAX_MOD_PARMS] = {0, };
+static int tx_coal_tick[ACE_MAX_MOD_PARMS] = {0, };
+static int rx_coal_tick[ACE_MAX_MOD_PARMS] = {0, };
+static int max_tx_desc[ACE_MAX_MOD_PARMS] = {0, };
+static int max_rx_desc[ACE_MAX_MOD_PARMS] = {0, };
+static int tx_ratio[ACE_MAX_MOD_PARMS] = {0, };
+static int dis_pci_mem_inval[ACE_MAX_MOD_PARMS] = {1, 1, 1, 1, 1, 1, 1, 1};
+
+static const char __initdata *version =
+ "acenic.c: v0.44 05/11/2000 Jes Sorensen, linux-acenic@SunSITE.auc.dk\n"
+ " http://home.cern.ch/~jes/gige/acenic.html\n";
+
+static struct net_device *root_dev = NULL;
static int probed __initdata = 0;
-__initfunc(int acenic_probe (struct device *dev))
+
+#ifdef NEW_NETINIT
+int __init acenic_probe (void)
+#else
+int __init acenic_probe (struct net_device *dev)
+#endif
{
- int boards_found = 0;
- int version_disp;
- struct ace_private *ap;
- u8 pci_latency;
-#if 0
- u16 vendor, device;
- u8 pci_bus;
- u8 pci_dev_fun;
- u8 irq;
+#ifdef NEW_NETINIT
+ struct net_device *dev;
#endif
+
+ struct ace_private *ap;
struct pci_dev *pdev = NULL;
+ int boards_found = 0;
+ int version_disp;
if (probed)
return -ENODEV;
@@ -210,11 +421,11 @@
version_disp = 0;
- while ((pdev = pci_find_class(PCI_CLASS_NETWORK_ETHERNET<<8, pdev))){
- dev = NULL;
+ while ((pdev = pci_find_class(PCI_CLASS_NETWORK_ETHERNET<<8, pdev))) {
if (!((pdev->vendor == PCI_VENDOR_ID_ALTEON) &&
- (pdev->device == PCI_DEVICE_ID_ALTEON_ACENIC)) &&
+ ((pdev->device == PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE) ||
+ (pdev->device == PCI_DEVICE_ID_ALTEON_ACENIC_COPPER)) ) &&
!((pdev->vendor == PCI_VENDOR_ID_3COM) &&
(pdev->device == PCI_DEVICE_ID_3COM_3C985)) &&
!((pdev->vendor == PCI_VENDOR_ID_NETGEAR) &&
@@ -229,44 +440,34 @@
(pdev->device == PCI_DEVICE_ID_SGI_ACENIC)))
continue;
- dev = init_etherdev(dev, sizeof(struct ace_private));
+ dev = init_etherdev(NULL, sizeof(struct ace_private));
- if (dev == NULL){
- printk(KERN_ERR "Unable to allocate etherdev "
- "structure!\n");
+ if (dev == NULL) {
+ printk(KERN_ERR "acenic: Unable to allocate "
+ "net_device structure!\n");
break;
}
if (!dev->priv)
dev->priv = kmalloc(sizeof(*ap), GFP_KERNEL);
- if (!dev->priv)
+ if (!dev->priv) {
+ printk(KERN_ERR "acenic: Unable to allocate memory\n");
return -ENOMEM;
+ }
ap = dev->priv;
ap->pdev = pdev;
- ap->vendor = pdev->vendor;
dev->irq = pdev->irq;
-#ifdef __SMP__
- spin_lock_init(&ap->lock);
-#endif
-
dev->open = &ace_open;
dev->hard_start_xmit = &ace_start_xmit;
dev->stop = &ace_close;
dev->get_stats = &ace_get_stats;
dev->set_multicast_list = &ace_set_multicast_list;
-#if 0
dev->do_ioctl = &ace_ioctl;
-#endif
dev->set_mac_address = &ace_set_mac_addr;
dev->change_mtu = &ace_change_mtu;
- /*
- * Dummy value.
- */
- dev->base_addr = 42;
-
/* display version info if adapter is found */
if (!version_disp)
{
@@ -278,76 +479,109 @@
pci_read_config_word(pdev, PCI_COMMAND, &ap->pci_command);
- pci_read_config_byte(pdev, PCI_LATENCY_TIMER, &pci_latency);
- if (pci_latency <= 0x40){
- pci_latency = 0x40;
+ /* OpenFirmware on Mac's does not set this - DOH.. */
+ if (!ap->pci_command & PCI_COMMAND_MEMORY) {
+ printk(KERN_INFO "%s: Enabling PCI Memory Mapped "
+ "access - was not enabled by BIOS/Firmware\n",
+ dev->name);
+ ap->pci_command = ap->pci_command | PCI_COMMAND_MEMORY;
+ pci_write_config_word(ap->pdev, PCI_COMMAND,
+ ap->pci_command);
+ wmb();
+ }
+
+ pci_read_config_byte(pdev, PCI_LATENCY_TIMER,
+ &ap->pci_latency);
+ if (ap->pci_latency <= 0x40) {
+ ap->pci_latency = 0x40;
pci_write_config_byte(pdev, PCI_LATENCY_TIMER,
- pci_latency);
+ ap->pci_latency);
}
pci_set_master(pdev);
- switch(ap->vendor){
+ /*
+ * Remap the regs into kernel space - this is abuse of
+ * dev->base_addr since it was means for I/O port
+ * addresses but who gives a damn.
+ */
+ dev->base_addr = pci_resource_start(pdev, 0);
+ ap->regs = (struct ace_regs *)ioremap(dev->base_addr, 0x4000);
+ if (!ap->regs) {
+ printk(KERN_ERR "%s: Unable to map I/O register, "
+ "AceNIC %i will be disabled.\n",
+ dev->name, boards_found);
+ break;
+ }
+
+ switch(pdev->vendor) {
case PCI_VENDOR_ID_ALTEON:
- sprintf(ap->name, "AceNIC Gigabit Ethernet");
+ strncpy(ap->name, "AceNIC Gigabit Ethernet",
+ sizeof (ap->name));
printk(KERN_INFO "%s: Alteon AceNIC ", dev->name);
break;
case PCI_VENDOR_ID_3COM:
- sprintf(ap->name, "3Com 3C985 Gigabit Ethernet");
+ strncpy(ap->name, "3Com 3C985 Gigabit Ethernet",
+ sizeof (ap->name));
printk(KERN_INFO "%s: 3Com 3C985 ", dev->name);
break;
case PCI_VENDOR_ID_NETGEAR:
- sprintf(ap->name, "NetGear GA620 Gigabit Ethernet");
+ strncpy(ap->name, "NetGear GA620 Gigabit Ethernet",
+ sizeof (ap->name));
printk(KERN_INFO "%s: NetGear GA620 ", dev->name);
break;
case PCI_VENDOR_ID_DEC:
if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9000SX) {
- sprintf(ap->name, "Farallon PN9000-SX "
- "Gigabit Ethernet");
+ strncpy(ap->name, "Farallon PN9000-SX "
+ "Gigabit Ethernet", sizeof (ap->name));
printk(KERN_INFO "%s: Farallon PN9000-SX ",
dev->name);
break;
}
case PCI_VENDOR_ID_SGI:
- sprintf(ap->name, "SGI AceNIC Gigabit Ethernet");
+ strncpy(ap->name, "SGI AceNIC Gigabit Ethernet",
+ sizeof (ap->name));
printk(KERN_INFO "%s: SGI AceNIC ", dev->name);
break;
default:
- sprintf(ap->name, "Unknown AceNIC based Gigabit Ethernet");
+ strncpy(ap->name, "Unknown AceNIC based Gigabit "
+ "Ethernet", sizeof (ap->name));
printk(KERN_INFO "%s: Unknown AceNIC ", dev->name);
break;
}
- printk("Gigabit Ethernet at 0x%08lx, irq %i, PCI latency %i "
- "clks\n", pdev->base_address[0], dev->irq, pci_latency);
-
- /*
- * Remap the regs into kernel space.
- */
+ ap->name [sizeof (ap->name) - 1] = '\0';
+ printk("Gigabit Ethernet at 0x%08lx, ", dev->base_addr);
+#ifdef __sparc__
+ printk("irq %s\n", __irq_itoa(dev->irq));
+#else
+ printk("irq %i\n", dev->irq);
+#endif
- ap->regs = (struct ace_regs *)ioremap(pdev->base_address[0],
- 0x4000);
- if (!ap->regs){
- printk(KERN_ERR "%s: Unable to map I/O register, "
- "AceNIC %i will be disabled.\n",
- dev->name, boards_found);
- break;
+#ifdef CONFIG_ACENIC_OMIT_TIGON_I
+ if ((readl(&ap->regs->HostCtrl) >> 28) == 4) {
+ printk(KERN_ERR "%s: Driver compiled without Tigon I"
+ " support - NIC disabled\n", dev->name);
+ ace_init_cleanup(dev);
+ continue;
}
+#endif
-#ifdef MODULE
- if (ace_init(dev, boards_found))
+ if (ace_allocate_descriptors(dev))
continue;
+
+#ifdef MODULE
+ if (boards_found >= ACE_MAX_MOD_PARMS)
+ ap->board_idx = BOARD_IDX_OVERFLOW;
+ else
+ ap->board_idx = boards_found;
#else
- if (ace_init(dev, -1))
- continue;
+ ap->board_idx = BOARD_IDX_STATIC;
#endif
- boards_found++;
+ if (ace_init(dev))
+ continue;
- /*
- * This is bollocks, but we need to tell the net-init
- * code that it shall go for the next device.
- */
- dev->base_addr = 0;
+ boards_found++;
}
/*
@@ -356,21 +590,16 @@
* or more boards. Otherwise, return failure (-ENODEV).
*/
-#ifdef MODULE
- return boards_found;
-#else
if (boards_found > 0)
return 0;
else
return -ENODEV;
-#endif
}
#ifdef MODULE
-#if LINUX_VERSION_CODE > 0x20118
MODULE_AUTHOR("Jes Sorensen <Jes.Sorensen@cern.ch>");
-MODULE_DESCRIPTION("AceNIC/3C985 Gigabit Ethernet driver");
+MODULE_DESCRIPTION("AceNIC/3C985/GA620 Gigabit Ethernet driver");
MODULE_PARM(link, "1-" __MODULE_STRING(8) "i");
MODULE_PARM(trace, "1-" __MODULE_STRING(8) "i");
MODULE_PARM(tx_coal_tick, "1-" __MODULE_STRING(8) "i");
@@ -379,64 +608,252 @@
MODULE_PARM(max_rx_desc, "1-" __MODULE_STRING(8) "i");
#endif
-int init_module(void)
-{
- int cards;
-
- root_dev = NULL;
-
- cards = acenic_probe(NULL);
- return cards ? 0 : -ENODEV;
-}
-void cleanup_module(void)
+void __exit ace_module_cleanup(void)
{
struct ace_private *ap;
struct ace_regs *regs;
- struct device *next;
+ struct net_device *next;
short i;
- unsigned long flags;
- while (root_dev){
+ while (root_dev) {
next = ((struct ace_private *)root_dev->priv)->next;
- ap = (struct ace_private *)root_dev->priv;
+ ap = root_dev->priv;
regs = ap->regs;
- spin_lock_irqsave(&ap->lock, flags);
writel(readl(®s->CpuCtrl) | CPU_HALT, ®s->CpuCtrl);
- if (ap->version == 2)
+ if (ap->version >= 2)
writel(readl(®s->CpuBCtrl) | CPU_HALT,
®s->CpuBCtrl);
- writel(0, ®s->Mb0Lo);
-
- spin_unlock_irqrestore(&ap->lock, flags);
+ /*
+ * This clears any pending interrupts
+ */
+ writel(1, ®s->Mb0Lo);
/*
- * Release the RX buffers.
+ * Make sure no other CPUs are processing interrupts
+ * on the card before the buffers are being released.
+ * Otherwise one might experience some `interesting'
+ * effects.
+ *
+ * Then release the RX buffers - jumbo buffers were
+ * already released in ace_close().
*/
+ synchronize_irq();
+
for (i = 0; i < RX_STD_RING_ENTRIES; i++) {
- if (ap->rx_std_skbuff[i]) {
+ struct sk_buff *skb = ap->skb->rx_std_skbuff[i].skb;
+
+ if (skb) {
+ dma_addr_t mapping;
+
+ mapping = ap->skb->rx_std_skbuff[i].mapping;
+
ap->rx_std_ring[i].size = 0;
- set_aceaddr_bus(&ap->rx_std_ring[i].addr, 0);
- dev_kfree_skb(ap->rx_std_skbuff[i]);
+ ap->skb->rx_std_skbuff[i].skb = NULL;
+ pci_unmap_single(ap->pdev, mapping,
+ ACE_STD_BUFSIZE - (2 + 16),
+ PCI_DMA_FROMDEVICE);
+ dev_kfree_skb(skb);
+ }
+ }
+ if (ap->version >= 2) {
+ for (i = 0; i < RX_MINI_RING_ENTRIES; i++) {
+ struct sk_buff *skb = ap->skb->rx_mini_skbuff[i].skb;
+
+ if (skb) {
+ dma_addr_t mapping;
+
+ mapping = ap->skb->rx_mini_skbuff[i].mapping;
+ ap->rx_mini_ring[i].size = 0;
+ ap->skb->rx_mini_skbuff[i].skb = NULL;
+ pci_unmap_single(ap->pdev, mapping,
+ ACE_MINI_BUFSIZE - (2 + 16),
+ PCI_DMA_FROMDEVICE);
+ dev_kfree_skb(skb);
+ }
}
}
+ for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
+ struct sk_buff *skb = ap->skb->rx_jumbo_skbuff[i].skb;
+ if (skb) {
+ dma_addr_t mapping;
- iounmap(regs);
- if(ap->trace_buf)
- kfree(ap->trace_buf);
- kfree(ap->info);
- free_irq(root_dev->irq, root_dev);
- unregister_netdev(root_dev);
- kfree(root_dev);
+ mapping = ap->skb->rx_jumbo_skbuff[i].mapping;
+ ap->rx_jumbo_ring[i].size = 0;
+ ap->skb->rx_jumbo_skbuff[i].skb = NULL;
+ pci_unmap_single(ap->pdev, mapping,
+ ACE_JUMBO_BUFSIZE - (2 + 16),
+ PCI_DMA_FROMDEVICE);
+ dev_kfree_skb(skb);
+ }
+ }
+
+ ace_init_cleanup(root_dev);
+ kfree(root_dev);
root_dev = next;
}
}
+
+
+int __init ace_module_init(void)
+{
+ int status;
+
+ root_dev = NULL;
+
+#ifdef NEW_NETINIT
+ status = acenic_probe();
+#else
+ status = acenic_probe(NULL);
+#endif
+ return status;
+}
+
+
+#if (LINUX_VERSION_CODE < 0x02032a)
+int init_module(void)
+{
+ return ace_module_init();
+}
+
+
+void cleanup_module(void)
+{
+ ace_module_cleanup();
+}
+#else
+module_init(ace_module_init);
+module_exit(ace_module_cleanup);
#endif
+static void ace_free_descriptors(struct net_device *dev)
+{
+ struct ace_private *ap = dev->priv;
+ int size;
+
+ if (ap->rx_std_ring != NULL) {
+ size = (sizeof(struct rx_desc) *
+ (RX_STD_RING_ENTRIES +
+ RX_JUMBO_RING_ENTRIES +
+ RX_MINI_RING_ENTRIES +
+ RX_RETURN_RING_ENTRIES));
+ pci_free_consistent(ap->pdev, size,
+ ap->rx_std_ring,
+ ap->rx_ring_base_dma);
+ ap->rx_std_ring = NULL;
+ ap->rx_jumbo_ring = NULL;
+ ap->rx_mini_ring = NULL;
+ ap->rx_return_ring = NULL;
+ }
+ if (ap->evt_ring != NULL) {
+ size = (sizeof(struct event) * EVT_RING_ENTRIES);
+ pci_free_consistent(ap->pdev, size,
+ ap->evt_ring,
+ ap->evt_ring_dma);
+ ap->evt_ring = NULL;
+ }
+ if (ap->evt_prd != NULL) {
+ pci_free_consistent(ap->pdev, sizeof(u32),
+ (void *)ap->evt_prd, ap->evt_prd_dma);
+ ap->evt_prd = NULL;
+ }
+ if (ap->rx_ret_prd != NULL) {
+ pci_free_consistent(ap->pdev, sizeof(u32),
+ (void *)ap->rx_ret_prd, ap->rx_ret_prd_dma);
+ ap->rx_ret_prd = NULL;
+ }
+ if (ap->tx_csm != NULL) {
+ pci_free_consistent(ap->pdev, sizeof(u32),
+ (void *)ap->tx_csm, ap->tx_csm_dma);
+ ap->tx_csm = NULL;
+ }
+}
+
+
+static int ace_allocate_descriptors(struct net_device *dev)
+{
+ struct ace_private *ap = dev->priv;
+ int size;
+
+ size = (sizeof(struct rx_desc) *
+ (RX_STD_RING_ENTRIES +
+ RX_JUMBO_RING_ENTRIES +
+ RX_MINI_RING_ENTRIES +
+ RX_RETURN_RING_ENTRIES));
+
+ ap->rx_std_ring = pci_alloc_consistent(ap->pdev, size,
+ &ap->rx_ring_base_dma);
+ if (ap->rx_std_ring == NULL)
+ goto fail;
+
+ ap->rx_jumbo_ring = ap->rx_std_ring + RX_STD_RING_ENTRIES;
+ ap->rx_mini_ring = ap->rx_jumbo_ring + RX_JUMBO_RING_ENTRIES;
+ ap->rx_return_ring = ap->rx_mini_ring + RX_MINI_RING_ENTRIES;
+
+ size = (sizeof(struct event) * EVT_RING_ENTRIES);
+
+ ap->evt_ring = pci_alloc_consistent(ap->pdev, size,
+ &ap->evt_ring_dma);
+
+ if (ap->evt_ring == NULL)
+ goto fail;
+
+ ap->evt_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
+ &ap->evt_prd_dma);
+ if (ap->evt_prd == NULL)
+ goto fail;
+
+ ap->rx_ret_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
+ &ap->rx_ret_prd_dma);
+ if (ap->rx_ret_prd == NULL)
+ goto fail;
+
+ ap->tx_csm = pci_alloc_consistent(ap->pdev, sizeof(u32),
+ &ap->tx_csm_dma);
+ if (ap->tx_csm == NULL)
+ goto fail;
+
+ return 0;
+
+fail:
+ /* Clean up. */
+ ace_init_cleanup(dev);
+ return 1;
+}
+
+
+/*
+ * Generic cleanup handling data allocated during init. Used when the
+ * module is unloaded or if an error occurs during initialization
+ */
+static void ace_init_cleanup(struct net_device *dev)
+{
+ struct ace_private *ap;
+
+ ap = dev->priv;
+
+ ace_free_descriptors(dev);
+
+ if (ap->info)
+ pci_free_consistent(ap->pdev, sizeof(struct ace_info),
+ ap->info, ap->info_dma);
+ if (ap->skb)
+ kfree(ap->skb);
+ if (ap->trace_buf)
+ kfree(ap->trace_buf);
+
+ if (dev->irq)
+ free_irq(dev->irq, dev);
+
+ unregister_netdev(dev);
+ iounmap(ap->regs);
+}
+
+
/*
* Commands are considered to be slow.
*/
@@ -453,24 +870,40 @@
}
-__initfunc(static int ace_init(struct device *dev, int board_idx))
+static int __init ace_init(struct net_device *dev)
{
struct ace_private *ap;
struct ace_regs *regs;
- struct ace_info *info;
- u32 tig_ver, mac1, mac2, tmp;
+ struct ace_info *info = NULL;
unsigned long tmp_ptr, myjif;
+ u32 tig_ver, mac1, mac2, tmp, pci_state;
+ int board_idx, ecode = 0;
short i;
+ unsigned char cache;
ap = dev->priv;
regs = ap->regs;
+ board_idx = ap->board_idx;
+
+ /*
+ * aman@sgi.com - its useful to do a NIC reset here to
+ * address the `Firmware not running' problem subsequent
+ * to any crashes involving the NIC
+ */
+ writel(HW_RESET | (HW_RESET << 24), ®s->HostCtrl);
+ wmb();
+
/*
* Don't access any other registes before this point!
*/
#ifdef __BIG_ENDIAN
- writel(((BYTE_SWAP | WORD_SWAP | CLR_INT) |
- ((BYTE_SWAP | WORD_SWAP | CLR_INT) << 24)),
+ /*
+ * This will most likely need BYTE_SWAP once we switch
+ * to using __raw_writel()
+ */
+ writel((WORD_SWAP | CLR_INT |
+ ((WORD_SWAP | CLR_INT) << 24)),
®s->HostCtrl);
#else
writel((CLR_INT | WORD_SWAP | ((CLR_INT | WORD_SWAP) << 24)),
@@ -487,18 +920,25 @@
tig_ver = readl(®s->HostCtrl) >> 28;
switch(tig_ver){
+#ifndef CONFIG_ACENIC_OMIT_TIGON_I
case 4:
- printk(KERN_INFO" Tigon I (Rev. 4), Firmware: %i.%i.%i, ",
+ printk(KERN_INFO" Tigon I (Rev. 4), Firmware: %i.%i.%i, ",
tigonFwReleaseMajor, tigonFwReleaseMinor,
tigonFwReleaseFix);
writel(0, ®s->LocalCtrl);
ap->version = 1;
break;
+#endif
case 6:
printk(KERN_INFO" Tigon II (Rev. %i), Firmware: %i.%i.%i, ",
tig_ver, tigon2FwReleaseMajor, tigon2FwReleaseMinor,
tigon2FwReleaseFix);
writel(readl(®s->CpuBCtrl) | CPU_HALT, ®s->CpuBCtrl);
+ /*
+ * The SRAM bank size does _not_ indicate the amount
+ * of memory on the card, it controls the _bank_ size!
+ * Ie. a 1MB AceNIC will have two banks of 512KB.
+ */
writel(SRAM_BANK_512K, ®s->LocalCtrl);
writel(SYNC_SRAM_TIMING, ®s->MiscCfg);
ap->version = 2;
@@ -506,7 +946,8 @@
default:
printk(KERN_INFO" Unsupported Tigon version detected (%i), ",
tig_ver);
- return -ENODEV;
+ ecode = -ENODEV;
+ goto init_error;
}
/*
@@ -516,22 +957,34 @@
* value a second time works as well. This is what caused the
* `Firmware not running' problem on the Tigon II.
*/
-#ifdef __LITTLE_ENDIAN
- writel(ACE_BYTE_SWAP_DATA | ACE_WARN | ACE_FATAL |
- ACE_WORD_SWAP | ACE_NO_JUMBO_FRAG, ®s->ModeStat);
+#ifdef __BIG_ENDIAN
+ writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL | ACE_BYTE_SWAP_BD |
+ ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, ®s->ModeStat);
#else
-#error "this driver doesn't run on big-endian machines yet!"
+ writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL |
+ ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, ®s->ModeStat);
#endif
+ mb();
mac1 = 0;
- for(i = 0; i < 4; i++){
+ for(i = 0; i < 4; i++) {
mac1 = mac1 << 8;
- mac1 |= read_eeprom_byte(regs, 0x8c+i);
+ tmp = read_eeprom_byte(dev, 0x8c+i);
+ if (tmp < 0) {
+ ecode = -EIO;
+ goto init_error;
+ } else
+ mac1 |= (tmp & 0xff);
}
mac2 = 0;
- for(i = 4; i < 8; i++){
+ for(i = 4; i < 8; i++) {
mac2 = mac2 << 8;
- mac2 |= read_eeprom_byte(regs, 0x8c+i);
+ tmp = read_eeprom_byte(dev, 0x8c+i);
+ if (tmp < 0) {
+ ecode = -EIO;
+ goto init_error;
+ } else
+ mac2 |= (tmp & 0xff);
}
writel(mac1, ®s->MacAddrHi);
@@ -549,24 +1002,57 @@
dev->dev_addr[5] = mac2 & 0xff;
/*
+ * Looks like this is necessary to deal with on all architectures,
+ * even this %$#%$# N440BX Intel based thing doesn't get it right.
+ * Ie. having two NICs in the machine, one will have the cache
+ * line set at boot time, the other will not.
+ */
+ pci_read_config_byte(ap->pdev, PCI_CACHE_LINE_SIZE, &cache);
+ if ((cache << 2) != SMP_CACHE_BYTES) {
+ printk(KERN_INFO " PCI cache line size set incorrectly "
+ "(%i bytes) by BIOS/FW, corring to %i\n",
+ (cache << 2), SMP_CACHE_BYTES);
+ pci_write_config_byte(ap->pdev, PCI_CACHE_LINE_SIZE,
+ SMP_CACHE_BYTES >> 2);
+ }
+
+ pci_state = readl(®s->PciState);
+ printk(KERN_INFO " PCI bus width: %i bits, speed: %iMHz, "
+ "latency: %i clks\n",
+ (pci_state & PCI_32BIT) ? 32 : 64,
+ (pci_state & PCI_66MHZ) ? 66 : 33,
+ ap->pci_latency);
+
+ /*
* Set the max DMA transfer size. Seems that for most systems
* the performance is better when no MAX parameter is
* set. However for systems enabling PCI write and invalidate,
* DMA writes must be set to the L1 cache line size to get
* optimal performance.
+ *
+ * The default is now to turn the PCI write and invalidate off
+ * - that is what Alteon does for NT.
*/
tmp = READ_CMD_MEM | WRITE_CMD_MEM;
- if (ap->version == 2){
-#if 0
+ if (ap->version >= 2) {
+ tmp |= (MEM_READ_MULTIPLE | (pci_state & PCI_66MHZ));
/*
- * According to the documentation this enables writes
- * to all PCI regs - NOT good.
+ * Tuning parameters only supported for 8 cards
*/
- tmp |= DMA_WRITE_ALL_ALIGN;
-#endif
- tmp |= MEM_READ_MULTIPLE;
- if (ap->pci_command & PCI_COMMAND_INVALIDATE){
- switch(L1_CACHE_BYTES){
+ if (board_idx == BOARD_IDX_OVERFLOW ||
+ dis_pci_mem_inval[board_idx]) {
+ if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
+ ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
+ pci_write_config_word(ap->pdev, PCI_COMMAND,
+ ap->pci_command);
+ printk(KERN_INFO " Disabling PCI memory "
+ "write and invalidate\n");
+ }
+ } else if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
+ printk(KERN_INFO " PCI memory write & invalidate "
+ "enabled by BIOS, enabling counter measures\n");
+
+ switch(SMP_CACHE_BYTES) {
case 16:
tmp |= DMA_WRITE_MAX_16;
break;
@@ -586,22 +1072,62 @@
}
}
}
+
+#ifdef __sparc__
+ /*
+ * On this platform, we know what the best dma settings
+ * are. We use 64-byte maximum bursts, because if we
+ * burst larger than the cache line size (or even cross
+ * a 64byte boundry in a single burst) the UltraSparc
+ * PCI controller will disconnect at 64-byte multiples.
+ *
+ * Read-multiple will be properly enabled above, and when
+ * set will give the PCI controller proper hints about
+ * prefetching.
+ */
+ tmp = tmp & ~DMA_READ_WRITE_MASK;
+ tmp |= DMA_READ_MAX_64;
+ tmp |= DMA_WRITE_MAX_64;
+#endif
writel(tmp, ®s->PciState);
- if (request_irq(dev->irq, ace_interrupt, SA_SHIRQ, ap->name, dev)) {
- printk(KERN_WARNING "%s: Requested IRQ %d is busy\n",
- dev->name, dev->irq);
- return -EAGAIN;
+#if 0
+ /*
+ * I have received reports from people having problems when this
+ * bit is enabled.
+ */
+ if (!(ap->pci_command & PCI_COMMAND_FAST_BACK)) {
+ printk(KERN_INFO " Enabling PCI Fast Back to Back\n");
+ ap->pci_command |= PCI_COMMAND_FAST_BACK;
+ pci_write_config_word(ap->pdev, PCI_COMMAND, ap->pci_command);
}
-
+#endif
+
/*
* Initialize the generic info block and the command+event rings
* and the control blocks for the transmit and receive rings
* as they need to be setup once and for all.
*/
- if (!(info = kmalloc(sizeof(struct ace_info), GFP_KERNEL))){
- free_irq(dev->irq, dev);
- return -EAGAIN;
+ if (!(info = pci_alloc_consistent(ap->pdev, sizeof(struct ace_info),
+ &ap->info_dma))) {
+ ecode = -EAGAIN;
+ goto init_error;
+ }
+ ap->info = info;
+
+ /*
+ * Get the memory for the skb rings.
+ */
+ if (!(ap->skb = kmalloc(sizeof(struct ace_skb), GFP_KERNEL))) {
+ ecode = -EAGAIN;
+ goto init_error;
+ }
+
+ if (request_irq(dev->irq, ace_interrupt, SA_SHIRQ, ap->name, dev)) {
+ printk(KERN_WARNING "%s: Requested IRQ %d is busy\n",
+ dev->name, dev->irq);
+ ecode = -EAGAIN;
+ goto init_error;
}
/*
@@ -611,13 +1137,20 @@
ap->next = root_dev;
root_dev = dev;
- ap->info = info;
- memset(info, 0, sizeof(struct ace_info));
+#ifdef INDEX_DEBUG
+ spin_lock_init(&ap->debug_lock);
+ ap->last_tx = TX_RING_ENTRIES - 1;
+ ap->last_std_rx = 0;
+ ap->last_mini_rx = 0;
+#endif
+
+ memset(ap->info, 0, sizeof(struct ace_info));
+ memset(ap->skb, 0, sizeof(struct ace_skb));
ace_load_firmware(dev);
ap->fw_running = 0;
- tmp_ptr = virt_to_bus((void *)info);
+ tmp_ptr = (unsigned long) ap->info_dma;
#if (BITS_PER_LONG == 64)
writel(tmp_ptr >> 32, ®s->InfoPtrHi);
#else
@@ -627,15 +1160,16 @@
memset(ap->evt_ring, 0, EVT_RING_ENTRIES * sizeof(struct event));
- set_aceaddr(&info->evt_ctrl.rngptr, ap->evt_ring);
+ set_aceaddr(&info->evt_ctrl.rngptr, ap->evt_ring_dma);
info->evt_ctrl.flags = 0;
- set_aceaddr(&info->evt_prd_ptr, &ap->evt_prd);
- ap->evt_prd = 0;
+ set_aceaddr(&info->evt_prd_ptr, ap->evt_prd_dma);
+ *(ap->evt_prd) = 0;
+ wmb();
writel(0, ®s->EvtCsm);
+ set_aceaddr(&info->cmd_ctrl.rngptr, 0x100);
info->cmd_ctrl.flags = 0;
- set_aceaddr_bus(&info->cmd_ctrl.rngptr, (void *)0x100);
info->cmd_ctrl.max_len = 0;
for (i = 0; i < CMD_RING_ENTRIES; i++)
@@ -644,55 +1178,92 @@
writel(0, ®s->CmdPrd);
writel(0, ®s->CmdCsm);
- set_aceaddr(&info->stats2_ptr, &info->s.stats);
+ tmp_ptr = ap->info_dma;
+ tmp_ptr += (unsigned long) &(((struct ace_info *)0)->s.stats);
+ set_aceaddr(&info->stats2_ptr, (dma_addr_t) tmp_ptr);
+ set_aceaddr(&info->rx_std_ctrl.rngptr, ap->rx_ring_base_dma);
info->rx_std_ctrl.max_len = ACE_STD_MTU + ETH_HLEN + 4;
- set_aceaddr(&info->rx_std_ctrl.rngptr, ap->rx_std_ring);
- info->rx_std_ctrl.flags = FLG_RX_TCP_UDP_SUM;
+ info->rx_std_ctrl.flags = RCB_FLG_TCP_UDP_SUM;
memset(ap->rx_std_ring, 0,
RX_STD_RING_ENTRIES * sizeof(struct rx_desc));
+ for (i = 0; i < RX_STD_RING_ENTRIES; i++)
+ ap->rx_std_ring[i].flags = BD_FLG_TCP_UDP_SUM;
+
+ ap->rx_std_skbprd = 0;
+ atomic_set(&ap->cur_rx_bufs, 0);
+
+ set_aceaddr(&info->rx_jumbo_ctrl.rngptr,
+ (ap->rx_ring_base_dma +
+ (sizeof(struct rx_desc) * RX_STD_RING_ENTRIES)));
info->rx_jumbo_ctrl.max_len = 0;
- set_aceaddr(&info->rx_jumbo_ctrl.rngptr, ap->rx_jumbo_ring);
- info->rx_jumbo_ctrl.flags = FLG_RX_TCP_UDP_SUM;
+ info->rx_jumbo_ctrl.flags = RCB_FLG_TCP_UDP_SUM;
memset(ap->rx_jumbo_ring, 0,
RX_JUMBO_RING_ENTRIES * sizeof(struct rx_desc));
- info->rx_mini_ctrl.max_len = 0;
-#if 0
- set_aceaddr(&info->rx_mini_ctrl.rngptr, ap->rx_mini_ring);
-#else
- set_aceaddr_bus(&info->rx_mini_ctrl.rngptr, 0);
-#endif
- info->rx_mini_ctrl.flags = FLG_RNG_DISABLED;
+ for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++)
+ ap->rx_jumbo_ring[i].flags = BD_FLG_TCP_UDP_SUM | BD_FLG_JUMBO;
+
+ ap->rx_jumbo_skbprd = 0;
+ atomic_set(&ap->cur_jumbo_bufs, 0);
-#if 0
memset(ap->rx_mini_ring, 0,
RX_MINI_RING_ENTRIES * sizeof(struct rx_desc));
-#endif
- set_aceaddr(&info->rx_return_ctrl.rngptr, ap->rx_return_ring);
+ if (ap->version >= 2) {
+ set_aceaddr(&info->rx_mini_ctrl.rngptr,
+ (ap->rx_ring_base_dma +
+ (sizeof(struct rx_desc) *
+ (RX_STD_RING_ENTRIES +
+ RX_JUMBO_RING_ENTRIES))));
+ info->rx_mini_ctrl.max_len = ACE_MINI_SIZE;
+ info->rx_mini_ctrl.flags = RCB_FLG_TCP_UDP_SUM;
+
+ for (i = 0; i < RX_MINI_RING_ENTRIES; i++)
+ ap->rx_mini_ring[i].flags =
+ BD_FLG_TCP_UDP_SUM | BD_FLG_MINI;
+ } else {
+ set_aceaddr(&info->rx_mini_ctrl.rngptr, 0);
+ info->rx_mini_ctrl.flags = RCB_FLG_RNG_DISABLE;
+ info->rx_mini_ctrl.max_len = 0;
+ }
+
+ ap->rx_mini_skbprd = 0;
+ atomic_set(&ap->cur_mini_bufs, 0);
+
+ set_aceaddr(&info->rx_return_ctrl.rngptr,
+ (ap->rx_ring_base_dma +
+ (sizeof(struct rx_desc) *
+ (RX_STD_RING_ENTRIES +
+ RX_JUMBO_RING_ENTRIES +
+ RX_MINI_RING_ENTRIES))));
info->rx_return_ctrl.flags = 0;
info->rx_return_ctrl.max_len = RX_RETURN_RING_ENTRIES;
memset(ap->rx_return_ring, 0,
RX_RETURN_RING_ENTRIES * sizeof(struct rx_desc));
- set_aceaddr(&info->rx_ret_prd_ptr, &ap->rx_ret_prd);
+ set_aceaddr(&info->rx_ret_prd_ptr, ap->rx_ret_prd_dma);
+ *(ap->rx_ret_prd) = 0;
writel(TX_RING_BASE, ®s->WinBase);
ap->tx_ring = (struct tx_desc *)regs->Window;
- for (i = 0; i < (TX_RING_ENTRIES * sizeof(struct tx_desc) / 4); i++){
+ for (i = 0; i < (TX_RING_ENTRIES * sizeof(struct tx_desc) / 4); i++) {
writel(0, (unsigned long)ap->tx_ring + i * 4);
}
+ set_aceaddr(&info->tx_ctrl.rngptr, TX_RING_BASE);
info->tx_ctrl.max_len = TX_RING_ENTRIES;
+#if TX_COAL_INTS_ONLY
+ info->tx_ctrl.flags = RCB_FLG_COAL_INT_ONLY;
+#else
info->tx_ctrl.flags = 0;
- set_aceaddr_bus(&info->tx_ctrl.rngptr, (void *)TX_RING_BASE);
+#endif
- set_aceaddr(&info->tx_csm_ptr, &ap->tx_csm);
+ set_aceaddr(&info->tx_csm_ptr, ap->tx_csm_dma);
/*
* Potential item for tuning parameter
@@ -705,21 +1276,15 @@
writel(1, ®s->AssistState);
writel(DEF_STAT, ®s->TuneStatTicks);
-
- writel(DEF_TX_COAL, ®s->TuneTxCoalTicks);
- writel(DEF_TX_MAX_DESC, ®s->TuneMaxTxDesc);
- writel(DEF_RX_COAL, ®s->TuneRxCoalTicks);
- writel(DEF_RX_MAX_DESC, ®s->TuneMaxRxDesc);
writel(DEF_TRACE, ®s->TuneTrace);
- writel(DEF_TX_RATIO, ®s->TxBufRat);
- if (board_idx >= 8) {
- printk(KERN_WARNING "%s: more then 8 NICs detected, "
- "ignoring module parameters!\n", dev->name);
- board_idx = -1;
- }
+ ace_set_rxtx_parms(dev, 0);
- if (board_idx >= 0) {
+ if (board_idx == BOARD_IDX_OVERFLOW) {
+ printk(KERN_WARNING "%s: more then %i NICs detected, "
+ "ignoring module parameters!\n",
+ dev->name, ACE_MAX_MOD_PARMS);
+ } else if (board_idx >= 0) {
if (tx_coal_tick[board_idx])
writel(tx_coal_tick[board_idx],
®s->TuneTxCoalTicks);
@@ -744,7 +1309,7 @@
*/
tmp = LNK_ENABLE | LNK_FULL_DUPLEX | LNK_1000MB | LNK_100MB |
LNK_10MB | LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL | LNK_NEGOTIATE;
- if(ap->version == 2)
+ if(ap->version >= 2)
tmp |= LNK_TX_FLOW_CTL_Y;
/*
@@ -755,7 +1320,7 @@
tmp = LNK_ENABLE;
- if (option & 0x01){
+ if (option & 0x01) {
printk(KERN_INFO "%s: Setting half duplex link\n",
dev->name);
tmp &= ~LNK_FULL_DUPLEX;
@@ -768,7 +1333,7 @@
tmp |= LNK_100MB;
if (option & 0x40)
tmp |= LNK_1000MB;
- if ((option & 0x70) == 0){
+ if ((option & 0x70) == 0) {
printk(KERN_WARNING "%s: No media speed specified, "
"forcing auto negotiation\n", dev->name);
tmp |= LNK_NEGOTIATE | LNK_1000MB |
@@ -781,28 +1346,42 @@
"negotiation\n", dev->name);
if (option & 0x200)
tmp |= LNK_RX_FLOW_CTL_Y;
- if ((option & 0x400) && (ap->version == 2)){
+ if ((option & 0x400) && (ap->version >= 2)) {
printk(KERN_INFO "%s: Enabling TX flow control\n",
dev->name);
tmp |= LNK_TX_FLOW_CTL_Y;
}
}
+ ap->link = tmp;
writel(tmp, ®s->TuneLink);
- if (ap->version == 2)
+ if (ap->version >= 2)
writel(tmp, ®s->TuneFastLink);
- if (ap->version == 1)
+ if (ACE_IS_TIGON_I(ap))
writel(tigonFwStartAddr, ®s->Pc);
- else if (ap->version == 2)
+ if (ap->version == 2)
writel(tigon2FwStartAddr, ®s->Pc);
writel(0, ®s->Mb0Lo);
/*
- * Start the NIC CPU
+ * Set tx_csm before we start receiving interrupts, otherwise
+ * the interrupt handler might think it is supposed to process
+ * tx ints before we are up and running, which may cause a null
+ * pointer access in the int handler.
*/
+ ap->tx_full = 0;
+ ap->cur_rx = 0;
+ ap->tx_prd = *(ap->tx_csm) = ap->tx_ret_csm = 0;
+ wmb();
+ ace_set_txprd(regs, ap, 0);
+ writel(0, ®s->RxRetCsm);
+
+ /*
+ * Start the NIC CPU
+ */
writel(readl(®s->CpuCtrl) & ~(CPU_HALT|CPU_TRACE), ®s->CpuCtrl);
/*
@@ -810,20 +1389,95 @@
*/
myjif = jiffies + 3 * HZ;
while (time_before(jiffies, myjif) && !ap->fw_running);
- if (!ap->fw_running){
+
+ if (!ap->fw_running) {
printk(KERN_ERR "%s: Firmware NOT running!\n", dev->name);
+
ace_dump_trace(ap);
writel(readl(®s->CpuCtrl) | CPU_HALT, ®s->CpuCtrl);
- return -EBUSY;
+
+ /* aman@sgi.com - account for badly behaving firmware/NIC:
+ * - have observed that the NIC may continue to generate
+ * interrupts for some reason; attempt to stop it - halt
+ * second CPU for Tigon II cards, and also clear Mb0
+ * - if we're a module, we'll fail to load if this was
+ * the only GbE card in the system => if the kernel does
+ * see an interrupt from the NIC, code to handle it is
+ * gone and OOps! - so free_irq also
+ */
+ if (ap->version >= 2)
+ writel(readl(®s->CpuBCtrl) | CPU_HALT,
+ ®s->CpuBCtrl);
+ writel(0, ®s->Mb0Lo);
+
+ ecode = -EBUSY;
+ goto init_error;
}
/*
* We load the ring here as there seem to be no way to tell the
* firmware to wipe the ring without re-initializing it.
*/
- ace_load_std_rx_ring(dev);
-
+ if (!test_and_set_bit(0, &ap->std_refill_busy))
+ ace_load_std_rx_ring(ap, RX_RING_SIZE);
+ else
+ printk(KERN_ERR "%s: Someone is busy refilling the RX ring\n",
+ dev->name);
+ if (ap->version >= 2) {
+ if (!test_and_set_bit(0, &ap->mini_refill_busy))
+ ace_load_mini_rx_ring(ap, RX_MINI_SIZE);
+ else
+ printk(KERN_ERR "%s: Someone is busy refilling "
+ "the RX mini ring\n", dev->name);
+ }
return 0;
+
+ init_error:
+ ace_init_cleanup(dev);
+ return ecode;
+}
+
+
+static void ace_set_rxtx_parms(struct net_device *dev, int jumbo)
+{
+ struct ace_private *ap;
+ struct ace_regs *regs;
+ int board_idx;
+
+ ap = dev->priv;
+ regs = ap->regs;
+
+ board_idx = ap->board_idx;
+
+ if (board_idx >= 0) {
+ if (!jumbo) {
+ if (!tx_coal_tick[board_idx])
+ writel(DEF_TX_COAL, ®s->TuneTxCoalTicks);
+ if (!max_tx_desc[board_idx])
+ writel(DEF_TX_MAX_DESC, ®s->TuneMaxTxDesc);
+ if (!rx_coal_tick[board_idx])
+ writel(DEF_RX_COAL, ®s->TuneRxCoalTicks);
+ if (!max_rx_desc[board_idx])
+ writel(DEF_RX_MAX_DESC, ®s->TuneMaxRxDesc);
+ if (!tx_ratio[board_idx])
+ writel(DEF_TX_RATIO, ®s->TxBufRat);
+ } else {
+ if (!tx_coal_tick[board_idx])
+ writel(DEF_JUMBO_TX_COAL,
+ ®s->TuneTxCoalTicks);
+ if (!max_tx_desc[board_idx])
+ writel(DEF_JUMBO_TX_MAX_DESC,
+ ®s->TuneMaxTxDesc);
+ if (!rx_coal_tick[board_idx])
+ writel(DEF_JUMBO_RX_COAL,
+ ®s->TuneRxCoalTicks);
+ if (!max_rx_desc[board_idx])
+ writel(DEF_JUMBO_RX_MAX_DESC,
+ ®s->TuneMaxRxDesc);
+ if (!tx_ratio[board_idx])
+ writel(DEF_JUMBO_TX_RATIO, ®s->TxBufRat);
+ }
+ }
}
@@ -832,8 +1486,8 @@
*/
static void ace_timer(unsigned long data)
{
- struct device *dev = (struct device *)data;
- struct ace_private *ap = (struct ace_private *)dev->priv;
+ struct net_device *dev = (struct net_device *)data;
+ struct ace_private *ap = dev->priv;
struct ace_regs *regs = ap->regs;
/*
@@ -841,7 +1495,7 @@
* seconds and there is data in the transmit queue, thus we
* asume the card is stuck.
*/
- if (ap->tx_csm != ap->tx_ret_csm){
+ if (*ap->tx_csm != ap->tx_ret_csm) {
printk(KERN_WARNING "%s: Transmitter is stuck, %08x\n",
dev->name, (unsigned int)readl(®s->HostCtrl));
}
@@ -851,6 +1505,44 @@
}
+static void ace_bh(struct net_device *dev)
+{
+ struct ace_private *ap = dev->priv;
+ int cur_size;
+
+ cur_size = atomic_read(&ap->cur_rx_bufs);
+ if ((cur_size < RX_LOW_STD_THRES) &&
+ !test_and_set_bit(0, &ap->std_refill_busy)) {
+#if DEBUG
+ printk("refilling buffers (current %i)\n", cur_size);
+#endif
+ ace_load_std_rx_ring(ap, RX_RING_SIZE - cur_size);
+ }
+
+ if (ap->version >= 2) {
+ cur_size = atomic_read(&ap->cur_mini_bufs);
+ if ((cur_size < RX_LOW_MINI_THRES) &&
+ !test_and_set_bit(0, &ap->mini_refill_busy)) {
+#if DEBUG
+ printk("refilling mini buffers (current %i)\n",
+ cur_size);
+#endif
+ ace_load_mini_rx_ring(ap, RX_MINI_SIZE - cur_size);
+ }
+ }
+
+ cur_size = atomic_read(&ap->cur_jumbo_bufs);
+ if (ap->jumbo && (cur_size < RX_LOW_JUMBO_THRES) &&
+ !test_and_set_bit(0, &ap->jumbo_refill_busy)) {
+#if DEBUG
+ printk("refilling jumbo buffers (current %i)\n", >cur_size);
+#endif
+ ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE - cur_size);
+ }
+ ap->bh_pending = 0;
+}
+
+
/*
* Copy the contents of the NIC's trace buffer to kernel memory.
*/
@@ -866,160 +1558,191 @@
/*
* Load the standard rx ring.
+ *
+ * Loading rings is safe without holding the spin lock since this is
+ * done only before the device is enabled, thus no interrupts are
+ * generated and by the interrupt handler/bh handler.
*/
-static int ace_load_std_rx_ring(struct device *dev)
+static void ace_load_std_rx_ring(struct ace_private *ap, int nr_bufs)
{
- struct ace_private *ap;
struct ace_regs *regs;
- struct ace_info *info;
- unsigned long flags;
- struct cmd cmd;
- short i;
+ short i, idx;
- ap = (struct ace_private *)dev->priv;
regs = ap->regs;
- info = ap->info;
- spin_lock_irqsave(&ap->lock, flags);
+ idx = ap->rx_std_skbprd;
- /*
- * Set tx_csm before we start receiving interrupts, otherwise
- * the interrupt handler might think it is supposed to process
- * tx ints before we are up and running, which may cause a null
- * pointer access in the int handler.
- */
- ap->tx_full = 0;
- ap->cur_rx = ap->dirty_rx = 0;
- ap->tx_prd = ap->tx_csm = ap->tx_ret_csm = 0;
- writel(0, ®s->RxRetCsm);
-
- for (i = 0; i < RX_RING_THRESH; i++) {
+ for (i = 0; i < nr_bufs; i++) {
struct sk_buff *skb;
+ struct rx_desc *rd;
+ dma_addr_t mapping;
- ap->rx_std_ring[i].flags = 0;
- skb = alloc_skb(ACE_STD_MTU + ETH_HLEN + 6, GFP_ATOMIC);
- ap->rx_std_skbuff[i] = skb;
+ skb = alloc_skb(ACE_STD_BUFSIZE, GFP_ATOMIC);
+ if (!skb)
+ break;
/*
- * Make sure the data contents end up on an aligned address
+ * Make sure IP header starts on a fresh cache line.
*/
- skb_reserve(skb, 2);
-
- set_aceaddr(&ap->rx_std_ring[i].addr, skb->data);
- ap->rx_std_ring[i].size = ACE_STD_MTU + ETH_HLEN + 4;
-
- ap->rx_std_ring[i].flags = 0;
- ap->rx_std_ring[i].type = DESC_RX;
-
- ap->rx_std_ring[i].idx = i;
- }
-
- ap->rx_std_skbprd = i;
-
- /*
- * The last descriptor needs to be marked as being special.
- */
- ap->rx_std_ring[i-1].type = DESC_END;
-
- cmd.evt = C_SET_RX_PRD_IDX;
- cmd.code = 0;
- cmd.idx = ap->rx_std_skbprd;
- ace_issue_cmd(regs, &cmd);
+ skb_reserve(skb, 2 + 16);
+ mapping = pci_map_single(ap->pdev, skb->data,
+ ACE_STD_BUFSIZE - (2 + 16),
+ PCI_DMA_FROMDEVICE);
+ ap->skb->rx_std_skbuff[idx].skb = skb;
+ ap->skb->rx_std_skbuff[idx].mapping = mapping;
+
+ rd = &ap->rx_std_ring[idx];
+ set_aceaddr(&rd->addr, mapping);
+ rd->size = ACE_STD_MTU + ETH_HLEN + 4;
+ rd->idx = idx;
+ idx = (idx + 1) % RX_STD_RING_ENTRIES;
+ }
+
+ if (!i)
+ goto error_out;
+
+ atomic_add(i, &ap->cur_rx_bufs);
+ ap->rx_std_skbprd = idx;
+
+ if (ACE_IS_TIGON_I(ap)) {
+ struct cmd cmd;
+ cmd.evt = C_SET_RX_PRD_IDX;
+ cmd.code = 0;
+ cmd.idx = ap->rx_std_skbprd;
+ ace_issue_cmd(regs, &cmd);
+ } else {
+ writel(idx, ®s->RxStdPrd);
+ wmb();
+ }
- spin_unlock_irqrestore(&ap->lock, flags);
+ out:
+ clear_bit(0, &ap->std_refill_busy);
+ return;
- return 0;
+ error_out:
+ printk(KERN_INFO "Out of memory when allocating "
+ "standard receive buffers\n");
+ goto out;
}
-/*
- * Load the jumbo rx ring, this may happen at any time if the MTU
- * is changed to a value > 1500.
- */
-static int ace_load_jumbo_rx_ring(struct device *dev)
+static void ace_load_mini_rx_ring(struct ace_private *ap, int nr_bufs)
{
- struct ace_private *ap;
struct ace_regs *regs;
- struct cmd cmd;
- unsigned long flags;
- short i;
+ short i, idx;
- ap = (struct ace_private *)dev->priv;
regs = ap->regs;
- spin_lock_irqsave(&ap->lock, flags);
-
- for (i = 0; i < RX_RING_JUMBO_THRESH; i++) {
+ idx = ap->rx_mini_skbprd;
+ for (i = 0; i < nr_bufs; i++) {
struct sk_buff *skb;
+ struct rx_desc *rd;
+ dma_addr_t mapping;
- ap->rx_jumbo_ring[i].flags = 0;
- skb = alloc_skb(ACE_JUMBO_MTU + ETH_HLEN + 6, GFP_ATOMIC);
- ap->rx_jumbo_skbuff[i] = skb;
+ skb = alloc_skb(ACE_MINI_BUFSIZE, GFP_ATOMIC);
+ if (!skb)
+ break;
/*
- * Make sure the data contents end up on an aligned address
+ * Make sure the IP header ends up on a fresh cache line
*/
- skb_reserve(skb, 2);
-
- set_aceaddr(&ap->rx_jumbo_ring[i].addr, skb->data);
- ap->rx_jumbo_ring[i].size = ACE_JUMBO_MTU + ETH_HLEN + 4;
+ skb_reserve(skb, 2 + 16);
+ mapping = pci_map_single(ap->pdev, skb->data,
+ ACE_MINI_BUFSIZE - (2 + 16),
+ PCI_DMA_FROMDEVICE);
+ ap->skb->rx_mini_skbuff[idx].skb = skb;
+ ap->skb->rx_mini_skbuff[idx].mapping = mapping;
- ap->rx_jumbo_ring[i].flags = DFLG_RX_JUMBO;
- ap->rx_jumbo_ring[i].type = DESC_RX;
-
- ap->rx_jumbo_ring[i].idx = i;
+ rd = &ap->rx_mini_ring[idx];
+ set_aceaddr(&rd->addr, mapping);
+ rd->size = ACE_MINI_SIZE;
+ rd->idx = idx;
+ idx = (idx + 1) % RX_MINI_RING_ENTRIES;
}
- ap->rx_jumbo_skbprd = i;
+ if (!i)
+ goto error_out;
- /*
- * The last descriptor needs to be marked as being special.
- */
- ap->rx_jumbo_ring[i-1].type = DESC_END;
+ atomic_add(i, &ap->cur_mini_bufs);
- cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
- cmd.code = 0;
- cmd.idx = ap->rx_jumbo_skbprd;
- ace_issue_cmd(regs, &cmd);
+ ap->rx_mini_skbprd = idx;
- spin_unlock_irqrestore(&ap->lock, flags);
+ writel(idx, ®s->RxMiniPrd);
+ wmb();
- return 0;
+ out:
+ clear_bit(0, &ap->mini_refill_busy);
+ return;
+ error_out:
+ printk(KERN_INFO "Out of memory when allocating "
+ "mini receive buffers\n");
+ goto out;
}
/*
- * Tell the firmware not to accept jumbos and flush the jumbo ring.
- * This function must be called with the spinlock held.
+ * Load the jumbo rx ring, this may happen at any time if the MTU
+ * is changed to a value > 1500.
*/
-static int ace_flush_jumbo_rx_ring(struct device *dev)
+static void ace_load_jumbo_rx_ring(struct ace_private *ap, int nr_bufs)
{
- struct ace_private *ap;
struct ace_regs *regs;
- struct cmd cmd;
- short i;
+ short i, idx;
- ap = (struct ace_private *)dev->priv;
regs = ap->regs;
- if (ap->jumbo){
- cmd.evt = C_RESET_JUMBO_RNG;
+ idx = ap->rx_jumbo_skbprd;
+
+ for (i = 0; i < nr_bufs; i++) {
+ struct sk_buff *skb;
+ struct rx_desc *rd;
+ dma_addr_t mapping;
+
+ skb = alloc_skb(ACE_JUMBO_BUFSIZE, GFP_ATOMIC);
+ if (!skb)
+ break;
+
+ /*
+ * Make sure the IP header ends up on a fresh cache line
+ */
+ skb_reserve(skb, 2 + 16);
+ mapping = pci_map_single(ap->pdev, skb->data,
+ ACE_JUMBO_BUFSIZE - (2 + 16),
+ PCI_DMA_FROMDEVICE);
+ ap->skb->rx_jumbo_skbuff[idx].skb = skb;
+ ap->skb->rx_jumbo_skbuff[idx].mapping = mapping;
+
+ rd = &ap->rx_jumbo_ring[idx];
+ set_aceaddr(&rd->addr, mapping);
+ rd->size = ACE_JUMBO_MTU + ETH_HLEN + 4;
+ rd->idx = idx;
+ idx = (idx + 1) % RX_JUMBO_RING_ENTRIES;
+ }
+
+ if (!i)
+ goto error_out;
+
+ atomic_add(i, &ap->cur_jumbo_bufs);
+ ap->rx_jumbo_skbprd = idx;
+
+ if (ACE_IS_TIGON_I(ap)) {
+ struct cmd cmd;
+ cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
cmd.code = 0;
- cmd.idx = 0;
+ cmd.idx = ap->rx_jumbo_skbprd;
ace_issue_cmd(regs, &cmd);
+ } else {
+ writel(idx, ®s->RxJumboPrd);
+ wmb();
+ }
- for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
- if (ap->rx_jumbo_skbuff[i]) {
- ap->rx_jumbo_ring[i].size = 0;
- set_aceaddr_bus(&ap->rx_jumbo_ring[i].addr, 0);
- dev_kfree_skb(ap->rx_jumbo_skbuff[i]);
- }
- }
- }else
- printk(KERN_ERR "%s: Trying to flush Jumbo ring without "
- "Jumbo support enabled\n", dev->name);
-
- return 0;
+ out:
+ clear_bit(0, &ap->jumbo_refill_busy);
+ return;
+ error_out:
+ printk(KERN_INFO "Out of memory when allocating "
+ "jumbo receive buffers\n");
+ goto out;
}
@@ -1028,37 +1751,46 @@
* events) and are handled here, outside the main interrupt handler,
* to reduce the size of the handler.
*/
-static u32 ace_handle_event(struct device *dev, u32 evtcsm, u32 evtprd)
+static u32 ace_handle_event(struct net_device *dev, u32 evtcsm, u32 evtprd)
{
struct ace_private *ap;
- ap = (struct ace_private *)dev->priv;
+ ap = dev->priv;
- while (evtcsm != evtprd){
- switch (ap->evt_ring[evtcsm].evt){
+ while (evtcsm != evtprd) {
+ switch (ap->evt_ring[evtcsm].evt) {
case E_FW_RUNNING:
printk(KERN_INFO "%s: Firmware up and running\n",
dev->name);
ap->fw_running = 1;
+ wmb();
break;
case E_STATS_UPDATED:
break;
case E_LNK_STATE:
{
u16 code = ap->evt_ring[evtcsm].code;
- if (code == E_C_LINK_UP){
- printk("%s: Optical link UP\n", dev->name);
- }
- else if (code == E_C_LINK_DOWN)
- printk(KERN_INFO "%s: Optical link DOWN\n",
+ switch (code) {
+ case E_C_LINK_UP:
+ printk(KERN_WARNING "%s: Optical link UP\n",
dev->name);
- else
- printk(KERN_INFO "%s: Unknown optical link "
+ break;
+ case E_C_LINK_DOWN:
+ printk(KERN_WARNING "%s: Optical link DOWN\n",
+ dev->name);
+ break;
+ case E_C_LINK_10_100:
+ printk(KERN_WARNING "%s: 10/100BaseT link "
+ "UP\n", dev->name);
+ break;
+ default:
+ printk(KERN_ERR "%s: Unknown optical link "
"state %02x\n", dev->name, code);
+ }
break;
}
case E_ERROR:
- switch(ap->evt_ring[evtcsm].code){
+ switch(ap->evt_ring[evtcsm].code) {
case E_C_ERR_INVAL_CMD:
printk(KERN_ERR "%s: invalid command error\n",
dev->name);
@@ -1077,7 +1809,24 @@
}
break;
case E_RESET_JUMBO_RNG:
+ {
+ int i;
+ for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
+ if (ap->skb->rx_jumbo_skbuff[i].skb) {
+ ap->rx_jumbo_ring[i].size = 0;
+ set_aceaddr(&ap->rx_jumbo_ring[i].addr, 0);
+ dev_kfree_skb(ap->skb->rx_jumbo_skbuff[i].skb);
+ ap->skb->rx_jumbo_skbuff[i].skb = NULL;
+ }
+ }
+ ap->jumbo = 0;
+ printk(KERN_INFO "%s: Jumbo ring flushed\n",
+ dev->name);
+ if (!ap->tx_full)
+ netif_wake_queue(dev);
+ clear_bit(0, &ap->jumbo_refill_busy);
break;
+ }
default:
printk(KERN_ERR "%s: Unhandled event 0x%02x\n",
dev->name, ap->evt_ring[evtcsm].evt);
@@ -1089,115 +1838,77 @@
}
-static int ace_rx_int(struct device *dev, u32 rxretprd, u32 rxretcsm)
+static void ace_rx_int(struct net_device *dev, u32 rxretprd, u32 rxretcsm)
{
- struct ace_private *ap = (struct ace_private *)dev->priv;
- struct ace_regs *regs = ap->regs;
- u32 idx, oldidx;
+ struct ace_private *ap = dev->priv;
+ u32 idx;
+ int mini_count = 0, std_count = 0;
idx = rxretcsm;
- while (idx != rxretprd){
- struct sk_buff *skb, *newskb, *oldskb;
- struct rx_desc *newrxdesc, *oldrxdesc;
- u32 prdidx, size;
- void *addr;
+ while (idx != rxretprd) {
+ struct ring_info *rip;
+ struct sk_buff *skb;
+ struct rx_desc *rxdesc, *retdesc;
+ u32 skbidx;
+ int desc_type, mapsize;
u16 csum;
- int jumbo;
-
- oldidx = ap->rx_return_ring[idx].idx;
- jumbo = ap->rx_return_ring[idx].flags & DFLG_RX_JUMBO;
-
- if (jumbo){
- oldskb = ap->rx_jumbo_skbuff[oldidx];
- prdidx = ap->rx_jumbo_skbprd;
- newrxdesc = &ap->rx_jumbo_ring[prdidx];
- oldrxdesc = &ap->rx_jumbo_ring[oldidx];
- }else{
- oldskb = ap->rx_std_skbuff[oldidx];
- prdidx = ap->rx_std_skbprd;
- newrxdesc = &ap->rx_std_ring[prdidx];
- oldrxdesc = &ap->rx_std_ring[oldidx];
- }
-
- size = oldrxdesc->size;
-
- if (size < PKT_COPY_THRESHOLD) {
- skb = alloc_skb(size + 2, GFP_ATOMIC);
- if (skb == NULL){
- printk(KERN_ERR "%s: Out of memory\n",
- dev->name);
- goto error;
- }
- /*
- * Make sure the real data is aligned
- */
- skb_reserve(skb, 2);
- memcpy(skb_put(skb, size), oldskb->data, size);
- addr = get_aceaddr_bus(&oldrxdesc->addr);
- newskb = oldskb;
- }else{
- skb = oldskb;
-
- skb_put(skb, size);
-
- newskb = alloc_skb(size + 2, GFP_ATOMIC);
- if (newskb == NULL){
- printk(KERN_ERR "%s: Out of memory\n",
- dev->name);
- goto error;
- }
+ retdesc = &ap->rx_return_ring[idx];
+ skbidx = retdesc->idx;
+ desc_type = retdesc->flags & (BD_FLG_JUMBO | BD_FLG_MINI);
+ switch(desc_type) {
/*
- * Make sure we DMA directly into nicely
- * aligned receive buffers
+ * Normal frames do not have any flags set
+ *
+ * Mini and normal frames arrive frequently,
+ * so use a local counter to avoid doing
+ * atomic operations for each packet arriving.
*/
- skb_reserve(newskb, 2);
- addr = (void *)virt_to_bus(newskb->data);
+ case 0:
+ rip = &ap->skb->rx_std_skbuff[skbidx];
+ mapsize = ACE_STD_BUFSIZE - (2 + 16);
+ rxdesc = &ap->rx_std_ring[skbidx];
+ std_count++;
+ break;
+ case BD_FLG_JUMBO:
+ rip = &ap->skb->rx_jumbo_skbuff[skbidx];
+ mapsize = ACE_JUMBO_BUFSIZE - (2 + 16);
+ rxdesc = &ap->rx_jumbo_ring[skbidx];
+ atomic_dec(&ap->cur_jumbo_bufs);
+ break;
+ case BD_FLG_MINI:
+ rip = &ap->skb->rx_mini_skbuff[skbidx];
+ mapsize = ACE_MINI_BUFSIZE - (2 + 16);
+ rxdesc = &ap->rx_mini_ring[skbidx];
+ mini_count++;
+ break;
+ default:
+ printk(KERN_INFO "%s: unknown frame type (0x%02x) "
+ "returned by NIC\n", dev->name,
+ retdesc->flags);
+ goto error;
}
- set_aceaddr_bus(&newrxdesc->addr, addr);
- newrxdesc->size = size;
-
- newrxdesc->flags = oldrxdesc->flags;
- newrxdesc->idx = prdidx;
- newrxdesc->type = DESC_RX;
-#if (BITS_PER_LONG == 32)
- newrxdesc->addr.addrhi = 0;
+ skb = rip->skb;
+ rip->skb = NULL;
+ pci_unmap_single(ap->pdev, rip->mapping, mapsize,
+ PCI_DMA_FROMDEVICE);
+ skb_put(skb, retdesc->size);
+#if 0
+ /* unncessary */
+ rxdesc->size = 0;
#endif
- oldrxdesc->size = 0;
- set_aceaddr_bus(&oldrxdesc->addr, 0);
-
- if (jumbo){
- ap->rx_jumbo_skbuff[oldidx] = NULL;
- ap->rx_jumbo_skbuff[prdidx] = newskb;
-
- prdidx = (prdidx + 1) % RX_JUMBO_RING_ENTRIES;
- ap->rx_jumbo_skbprd = prdidx;
- }else{
- ap->rx_std_skbuff[oldidx] = NULL;
- ap->rx_std_skbuff[prdidx] = newskb;
-
- prdidx = (prdidx + 1) % RX_STD_RING_ENTRIES;
- ap->rx_std_skbprd = prdidx;
- }
-
/*
* Fly baby, fly!
*/
- csum = ap->rx_return_ring[idx].tcp_udp_csum;
+ csum = retdesc->tcp_udp_csum;
skb->dev = dev;
skb->protocol = eth_type_trans(skb, dev);
-#if 0
- /*
- * This was never actually enabled in the RX descriptors
- * anyway - it requires a bit more testing before enabling
- * it again.
- */
/*
* If the checksum is correct and this is not a
* fragment, tell the stack that the data is correct.
@@ -1208,34 +1919,31 @@
skb->ip_summed = CHECKSUM_UNNECESSARY;
else
skb->ip_summed = CHECKSUM_NONE;
-#endif
+
netif_rx(skb); /* send it up */
ap->stats.rx_packets++;
- ap->stats.rx_bytes += skb->len;
-
- if ((prdidx & 0x7) == 0){
- struct cmd cmd;
- if (jumbo)
- cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
- else
- cmd.evt = C_SET_RX_PRD_IDX;
- cmd.code = 0;
- cmd.idx = prdidx;
- ace_issue_cmd(regs, &cmd);
- }
+ ap->stats.rx_bytes += retdesc->size;
idx = (idx + 1) % RX_RETURN_RING_ENTRIES;
}
+
+ atomic_sub(std_count, &ap->cur_rx_bufs);
+ if (!ACE_IS_TIGON_I(ap))
+ atomic_sub(mini_count, &ap->cur_mini_bufs);
+
out:
/*
* According to the documentation RxRetCsm is obsolete with
- * the 12.3.x Firmware - my Tigon I NIC's seem to disagree!
+ * the 12.3.x Firmware - my Tigon I NICs seem to disagree!
*/
- writel(idx, ®s->RxRetCsm);
+ if (ACE_IS_TIGON_I(ap)) {
+ struct ace_regs *regs = ap->regs;
+ writel(idx, ®s->RxRetCsm);
+ }
ap->cur_rx = idx;
- return idx;
+ return;
error:
idx = rxretprd;
goto out;
@@ -1246,24 +1954,21 @@
{
struct ace_private *ap;
struct ace_regs *regs;
- struct device *dev = (struct device *)dev_id;
+ struct net_device *dev = (struct net_device *)dev_id;
+ u32 idx;
u32 txcsm, rxretcsm, rxretprd;
u32 evtcsm, evtprd;
- ap = (struct ace_private *)dev->priv;
+ ap = dev->priv;
regs = ap->regs;
- spin_lock(&ap->lock);
-
/*
* In case of PCI shared interrupts or spurious interrupts,
* we want to make sure it is actually our interrupt before
* spending any time in here.
*/
- if (!(readl(®s->HostCtrl) & IN_INT)){
- spin_unlock(&ap->lock);
+ if (!(readl(®s->HostCtrl) & IN_INT))
return;
- }
/*
* Tell the card not to generate interrupts while we are in here.
@@ -1271,25 +1976,46 @@
writel(1, ®s->Mb0Lo);
/*
- * Service RX ints before TX
+ * There is no conflict between transmit handling in
+ * start_xmit and receive processing, thus there is no reason
+ * to take a spin lock for RX handling. Wait until we start
+ * working on the other stuff - hey we don't need a spin lock
+ * anymore.
*/
- rxretprd = ap->rx_ret_prd;
+ rxretprd = *ap->rx_ret_prd;
rxretcsm = ap->cur_rx;
if (rxretprd != rxretcsm)
- rxretprd = ace_rx_int(dev, rxretprd, rxretcsm);
+ ace_rx_int(dev, rxretprd, rxretcsm);
- txcsm = ap->tx_csm;
- if (txcsm != ap->tx_ret_csm) {
- u32 idx = ap->tx_ret_csm;
+ txcsm = *ap->tx_csm;
+ idx = ap->tx_ret_csm;
+ if (txcsm != idx) {
do {
- ap->stats.tx_packets++;
- ap->stats.tx_bytes += ap->tx_skbuff[idx]->len;
- dev_kfree_skb(ap->tx_skbuff[idx]);
+ struct sk_buff *skb;
+
+ skb = ap->skb->tx_skbuff[idx].skb;
+ if (skb) {
+ dma_addr_t mapping;
- ap->tx_skbuff[idx] = NULL;
+ mapping = ap->skb->tx_skbuff[idx].mapping;
+ ap->stats.tx_packets++;
+ ap->stats.tx_bytes += skb->len;
+ pci_unmap_single(ap->pdev, mapping, skb->len,
+ PCI_DMA_TODEVICE);
+ dev_kfree_skb_irq(skb);
+
+ ap->skb->tx_skbuff[idx].skb = NULL;
+ }
+
+ /*
+ * Question here is whether one should not skip
+ * these writes - I have never seen any errors
+ * caused by the NIC actually trying to access
+ * these incorrectly.
+ */
#if (BITS_PER_LONG == 64)
writel(0, &ap->tx_ring[idx].addr.addrhi);
#endif
@@ -1299,11 +2025,19 @@
idx = (idx + 1) % TX_RING_ENTRIES;
} while (idx != txcsm);
- if (ap->tx_full && dev->tbusy &&
- (((ap->tx_prd + 1) % TX_RING_ENTRIES) != txcsm)){
- ap->tx_full = 0;
- dev->tbusy = 0;
- mark_bh(NET_BH);
+ /*
+ * Once we actually get to this point the tx ring has
+ * already been trimmed thus it cannot be full!
+ * Ie. skip the comparison of the tx producer vs. the
+ * consumer.
+ */
+ if (netif_queue_stopped(dev) && xchg(&ap->tx_full, 0)) {
+ /*
+ * This does not need to be atomic (and expensive),
+ * I've seen cases where it would fail otherwise ;-(
+ */
+ netif_wake_queue(dev);
+ ace_mark_net_bh(NET_BH);
/*
* TX ring is no longer full, aka the
@@ -1313,23 +2047,84 @@
}
ap->tx_ret_csm = txcsm;
+ wmb();
}
evtcsm = readl(®s->EvtCsm);
- evtprd = ap->evt_prd;
+ evtprd = *ap->evt_prd;
- if (evtcsm != evtprd){
+ if (evtcsm != evtprd) {
evtcsm = ace_handle_event(dev, evtcsm, evtprd);
+ writel(evtcsm, ®s->EvtCsm);
}
- writel(evtcsm, ®s->EvtCsm);
- writel(0, ®s->Mb0Lo);
+ /*
+ * This has to go last in the interrupt handler and run with
+ * the spin lock released ... what lock?
+ */
+ if (netif_running(dev)) {
+ int cur_size;
+ int run_bh = 0;
+
+ cur_size = atomic_read(&ap->cur_rx_bufs);
+ if (cur_size < RX_LOW_STD_THRES) {
+ if ((cur_size < RX_PANIC_STD_THRES) &&
+ !test_and_set_bit(0, &ap->std_refill_busy)) {
+#if DEBUG
+ printk("low on std buffers %i\n", cur_size);
+#endif
+ ace_load_std_rx_ring(ap,
+ RX_RING_SIZE - cur_size);
+ } else
+ run_bh = 1;
+ }
+
+ if (!ACE_IS_TIGON_I(ap)) {
+ cur_size = atomic_read(&ap->cur_mini_bufs);
+ if (cur_size < RX_LOW_MINI_THRES) {
+ if ((cur_size < RX_PANIC_MINI_THRES) &&
+ !test_and_set_bit(0,
+ &ap->mini_refill_busy)) {
+#if DEBUG
+ printk("low on mini buffers %i\n",
+ cur_size);
+#endif
+ ace_load_mini_rx_ring(ap, RX_MINI_SIZE - cur_size);
+ } else
+ run_bh = 1;
+ }
+ }
+
+ if (ap->jumbo) {
+ cur_size = atomic_read(&ap->cur_jumbo_bufs);
+ if (cur_size < RX_LOW_JUMBO_THRES) {
+ if ((cur_size < RX_PANIC_JUMBO_THRES) &&
+ !test_and_set_bit(0,
+ &ap->jumbo_refill_busy)){
+#if DEBUG
+ printk("low on jumbo buffers %i\n",
+ cur_size);
+#endif
+ ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE - cur_size);
+ } else
+ run_bh = 1;
+ }
+ }
+ if (run_bh && !ap->bh_pending) {
+ ap->bh_pending = 1;
+ queue_task(&ap->immediate, &tq_immediate);
+ mark_bh(IMMEDIATE_BH);
+ }
+ }
- spin_unlock(&ap->lock);
+ /*
+ * Allow the card to generate interrupts again
+ */
+ writel(0, ®s->Mb0Lo);
}
-static int ace_open(struct device *dev)
+static int ace_open(struct net_device *dev)
{
struct ace_private *ap;
struct ace_regs *regs;
@@ -1338,22 +2133,33 @@
ap = dev->priv;
regs = ap->regs;
- if (!(ap->fw_running)){
- printk(KERN_WARNING "%s: firmware not running!\n", dev->name);
+ if (!(ap->fw_running)) {
+ printk(KERN_WARNING "%s: Firmware not running!\n", dev->name);
return -EBUSY;
}
writel(dev->mtu + ETH_HLEN + 4, ®s->IfMtu);
+ /*
+ * Zero the stats when restarting the interface...
+ */
+ memset(&ap->stats, 0, sizeof(ap->stats));
+
+ cmd.evt = C_CLEAR_STATS;
+ cmd.code = 0;
+ cmd.idx = 0;
+ ace_issue_cmd(regs, &cmd);
+
cmd.evt = C_HOST_STATE;
cmd.code = C_C_STACK_UP;
cmd.idx = 0;
ace_issue_cmd(regs, &cmd);
- if (ap->jumbo)
- ace_load_jumbo_rx_ring(dev);
+ if (ap->jumbo &&
+ !test_and_set_bit(0, &ap->jumbo_refill_busy))
+ ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE);
- if (dev->flags & IFF_PROMISC){
+ if (dev->flags & IFF_PROMISC) {
cmd.evt = C_SET_PROMISC_MODE;
cmd.code = C_C_PROMISC_ENABLE;
cmd.idx = 0;
@@ -1365,19 +2171,13 @@
ap->mcast_all = 0;
#if 0
- { long myjif = jiffies + HZ;
- while (time_before(jiffies, myjif));
- }
-
cmd.evt = C_LNK_NEGOTIATION;
cmd.code = 0;
cmd.idx = 0;
ace_issue_cmd(regs, &cmd);
#endif
- dev->tbusy = 0;
- dev->interrupt = 0;
- dev->start = 1;
+ netif_start_queue(dev);
MOD_INC_USE_COUNT;
@@ -1387,11 +2187,20 @@
init_timer(&ap->timer);
ap->timer.data = (unsigned long)dev;
ap->timer.function = ace_timer;
+
+ /*
+ * Setup the bottom half rx ring refill handler
+ */
+ ap->immediate.next = NULL;
+ ap->immediate.sync = 0;
+ ap->immediate.routine = (void *)(void *)ace_bh;
+ ap->immediate.data = dev;
+
return 0;
}
-static int ace_close(struct device *dev)
+static int ace_close(struct net_device *dev)
{
struct ace_private *ap;
struct ace_regs *regs;
@@ -1399,15 +2208,15 @@
unsigned long flags;
short i;
- dev->start = 0;
- set_bit(0, (void*)&dev->tbusy);
+ ace_if_down(dev);
+ netif_stop_queue(dev);
- ap = (struct ace_private *)dev->priv;
+ ap = dev->priv;
regs = ap->regs;
del_timer(&ap->timer);
- if (ap->promisc){
+ if (ap->promisc) {
cmd.evt = C_SET_PROMISC_MODE;
cmd.code = C_C_PROMISC_DISABLE;
cmd.idx = 0;
@@ -1420,56 +2229,95 @@
cmd.idx = 0;
ace_issue_cmd(regs, &cmd);
- spin_lock_irqsave(&ap->lock, flags);
+ /*
+ * Make sure one CPU is not processing packets while
+ * buffers are being released by another.
+ */
+ save_flags(flags);
+ cli();
for (i = 0; i < TX_RING_ENTRIES; i++) {
- if (ap->tx_skbuff[i]) {
+ struct sk_buff *skb;
+ dma_addr_t mapping;
+
+ skb = ap->skb->tx_skbuff[i].skb;
+ mapping = ap->skb->tx_skbuff[i].mapping;
+ if (skb) {
writel(0, &ap->tx_ring[i].addr.addrhi);
writel(0, &ap->tx_ring[i].addr.addrlo);
writel(0, &ap->tx_ring[i].flagsize);
- dev_kfree_skb(ap->tx_skbuff[i]);
+ pci_unmap_single(ap->pdev, mapping, skb->len,
+ PCI_DMA_TODEVICE);
+ dev_kfree_skb(skb);
+ ap->skb->tx_skbuff[i].skb = NULL;
}
}
- if (ap->jumbo)
- ace_flush_jumbo_rx_ring(dev);
+ if (ap->jumbo) {
+ cmd.evt = C_RESET_JUMBO_RNG;
+ cmd.code = 0;
+ cmd.idx = 0;
+ ace_issue_cmd(regs, &cmd);
+ }
- spin_unlock_irqrestore(&ap->lock, flags);
+ restore_flags(flags);
MOD_DEC_USE_COUNT;
return 0;
}
-static int ace_start_xmit(struct sk_buff *skb, struct device *dev)
+static int ace_start_xmit(struct sk_buff *skb, struct net_device *dev)
{
- struct ace_private *ap = (struct ace_private *)dev->priv;
+ struct ace_private *ap = dev->priv;
struct ace_regs *regs = ap->regs;
- unsigned long flags;
unsigned long addr;
u32 idx, flagsize;
- spin_lock_irqsave(&ap->lock, flags);
+ /*
+ * ARGH, there is just no pretty way to do this
+ */
+#if (LINUX_VERSION_CODE < 0x02032b)
+ if (test_and_set_bit(0, &dev->tbusy))
+ return 1;
+#else
+ netif_stop_queue(dev);
+#endif
idx = ap->tx_prd;
- ap->tx_skbuff[idx] = skb;
- addr = virt_to_bus(skb->data);
+ if ((idx + 1) % TX_RING_ENTRIES == ap->tx_ret_csm) {
+ ap->tx_full = 1;
+#if DEBUG
+ printk("%s: trying to transmit while the tx ring is full "
+ "- this should not happen!\n", dev->name);
+#endif
+ return 1;
+ }
+
+ ap->skb->tx_skbuff[idx].skb = skb;
+ ap->skb->tx_skbuff[idx].mapping =
+ pci_map_single(ap->pdev, skb->data, skb->len,
+ PCI_DMA_TODEVICE);
+ addr = (unsigned long) ap->skb->tx_skbuff[idx].mapping;
#if (BITS_PER_LONG == 64)
writel(addr >> 32, &ap->tx_ring[idx].addr.addrhi);
#endif
writel(addr & 0xffffffff, &ap->tx_ring[idx].addr.addrlo);
- flagsize = (skb->len << 16) | (DESC_END) ;
+ flagsize = (skb->len << 16) | (BD_FLG_END) ;
writel(flagsize, &ap->tx_ring[idx].flagsize);
- mb();
+ wmb();
idx = (idx + 1) % TX_RING_ENTRIES;
ap->tx_prd = idx;
- writel(idx, ®s->TxPrd);
+ ace_set_txprd(regs, ap, idx);
- if ((idx + 1) % TX_RING_ENTRIES == ap->tx_ret_csm){
+ /*
+ * tx_csm is set by the NIC whereas we set tx_ret_csm which
+ * is always trying to catch tx_csm
+ */
+ if ((idx + 2) % TX_RING_ENTRIES == ap->tx_ret_csm) {
ap->tx_full = 1;
- set_bit(0, (void*)&dev->tbusy);
/*
* Queue is full, add timer to detect whether the
* transmitter is stuck. Use mod_timer as we can get
@@ -1477,16 +2325,33 @@
* timers.
*/
mod_timer(&ap->timer, jiffies + (3 * HZ));
- }
- spin_unlock_irqrestore(&ap->lock, flags);
+ /* The following check will fix a race between the interrupt
+ * handler increasing the tx_ret_csm and testing for tx_full
+ * and this tx routine's testing the tx_ret_csm and setting
+ * the tx_full; note that this fix makes assumptions on the
+ * ordering of writes (sequential consistency will fly; TSO
+ * processor order would work too) but that's what lock-less
+ * programming is all about
+ */
+ if (((idx + 2) % TX_RING_ENTRIES != ap->tx_ret_csm)
+ && xchg(&ap->tx_full, 0)) {
+ del_timer(&ap->timer);
+ netif_wake_queue(dev);
+ }
+ } else {
+ /*
+ * No need for it to be atomic - seems it needs to be
+ */
+ netif_wake_queue(dev);
+ }
dev->trans_start = jiffies;
return 0;
}
-static int ace_change_mtu(struct device *dev, int new_mtu)
+static int ace_change_mtu(struct net_device *dev, int new_mtu)
{
struct ace_private *ap = dev->priv;
struct ace_regs *regs = ap->regs;
@@ -1497,40 +2362,182 @@
writel(new_mtu + ETH_HLEN + 4, ®s->IfMtu);
dev->mtu = new_mtu;
- if (new_mtu > ACE_STD_MTU){
- if (!(ap->jumbo)){
+ if (new_mtu > ACE_STD_MTU) {
+ if (!(ap->jumbo)) {
printk(KERN_INFO "%s: Enabling Jumbo frame "
"support\n", dev->name);
ap->jumbo = 1;
- ace_load_jumbo_rx_ring(dev);
- }
- ap->jumbo = 1;
- }else{
+ if (!test_and_set_bit(0, &ap->jumbo_refill_busy))
+ ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE);
+ ace_set_rxtx_parms(dev, 1);
+ }
+ } else {
+ netif_stop_queue(dev);
+ while (test_and_set_bit(0, &ap->jumbo_refill_busy));
+ synchronize_irq();
+ ace_set_rxtx_parms(dev, 0);
if (ap->jumbo){
- ace_flush_jumbo_rx_ring(dev);
+ struct cmd cmd;
- printk(KERN_INFO "%s: Disabling Jumbo frame support\n",
- dev->name);
+ cmd.evt = C_RESET_JUMBO_RNG;
+ cmd.code = 0;
+ cmd.idx = 0;
+ ace_issue_cmd(regs, &cmd);
}
- ap->jumbo = 0;
}
return 0;
}
+static int ace_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
+{
+#ifdef ETHTOOL
+ struct ace_private *ap = dev->priv;
+ struct ace_regs *regs = ap->regs;
+ struct ethtool_cmd ecmd;
+ u32 link, speed;
+
+#ifdef SPIN_DEBUG
+ if (cmd == (SIOCDEVPRIVATE+0x0e)) {
+ printk(KERN_NOTICE "%s: dumping debug info\n", dev->name);
+ printk(KERN_NOTICE "%s: tbusy %li, tx_ret_csm %i, "
+ "tx_prd %i\n", dev->name, dev->tbusy,
+ ap->tx_ret_csm, ap->tx_prd);
+ printk(KERN_NOTICE "%s: cur_rx %i, std_refill %li, "
+ "mini_rx %i, mini_refill %li\n", dev->name,
+ atomic_read(&ap->cur_rx_bufs), ap->std_refill_busy,
+ atomic_read(&ap->cur_mini_bufs), ap->mini_refill_busy);
+ printk(KERN_NOTICE "%s: CpuCtrl %08x\n",
+ dev->name, readl(®s->CpuCtrl));
+ return 0;
+ }
+#endif
+ if (cmd != SIOCETHTOOL)
+ return -EOPNOTSUPP;
+ if (copy_from_user(&ecmd, ifr->ifr_data, sizeof(ecmd)))
+ return -EFAULT;
+
+ if (ecmd.cmd == ETH_GSET) {
+ ecmd.supported =
+ (SUPPORTED_10baseT_Half | SUPPORTED_10baseT_Full |
+ SUPPORTED_100baseT_Half | SUPPORTED_100baseT_Full |
+ SUPPORTED_1000baseT_Half | SUPPORTED_1000baseT_Full |
+ SUPPORTED_Autoneg | SUPPORTED_FIBRE);
+
+ ecmd.port = PORT_FIBRE;
+ ecmd.transceiver = XCVR_INTERNAL;
+ ecmd.phy_address = 0;
+
+ link = readl(®s->GigLnkState);
+ if (link & LNK_1000MB)
+ ecmd.speed = SPEED_1000;
+ else {
+ link = readl(®s->FastLnkState);
+ if (link & LNK_100MB)
+ ecmd.speed = SPEED_100;
+ else if (link & LNK_100MB)
+ ecmd.speed = SPEED_10;
+ else
+ ecmd.speed = 0;
+ }
+ if (link & LNK_FULL_DUPLEX)
+ ecmd.duplex = DUPLEX_FULL;
+ else
+ ecmd.duplex = DUPLEX_HALF;
+
+ if (link & LNK_NEGOTIATE)
+ ecmd.autoneg = AUTONEG_ENABLE;
+ else
+ ecmd.autoneg = AUTONEG_DISABLE;
+
+ ecmd.trace = readl(®s->TuneTrace);
+
+ ecmd.txcoal = readl(®s->TuneTxCoalTicks);
+ ecmd.rxcoal = readl(®s->TuneRxCoalTicks);
+ ecmd.maxtxpkt = readl(®s->TuneMaxTxDesc);
+ ecmd.maxrxpkt = readl(®s->TuneMaxRxDesc);
+
+ if(copy_to_user(ifr->ifr_data, &ecmd, sizeof(ecmd)))
+ return -EFAULT;
+ return 0;
+ } else if (ecmd.cmd == ETH_SSET) {
+ if(!capable(CAP_NET_ADMIN))
+ return -EPERM;
+
+ link = readl(®s->GigLnkState);
+ if (link & LNK_1000MB)
+ speed = SPEED_1000;
+ else {
+ link = readl(®s->FastLnkState);
+ if (link & LNK_100MB)
+ speed = SPEED_100;
+ else if (link & LNK_100MB)
+ speed = SPEED_10;
+ else
+ speed = SPEED_100;
+ }
+
+ link = LNK_ENABLE | LNK_1000MB | LNK_100MB | LNK_10MB |
+ LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL;
+ if (!ACE_IS_TIGON_I(ap))
+ link |= LNK_TX_FLOW_CTL_Y;
+ if (ecmd.autoneg == AUTONEG_ENABLE)
+ link |= LNK_NEGOTIATE;
+ if (ecmd.speed != speed) {
+ link &= ~(LNK_1000MB | LNK_100MB | LNK_10MB);
+ switch (speed) {
+ case SPEED_1000:
+ link |= LNK_1000MB;
+ break;
+ case SPEED_100:
+ link |= LNK_100MB;
+ break;
+ case SPEED_10:
+ link |= LNK_10MB;
+ break;
+ }
+ }
+ if (ecmd.duplex == DUPLEX_FULL)
+ link |= LNK_FULL_DUPLEX;
+
+ if (link != ap->link) {
+ struct cmd cmd;
+ printk(KERN_INFO "%s: Renegotiating link state\n",
+ dev->name);
+
+ ap->link = link;
+ writel(link, ®s->TuneLink);
+ if (!ACE_IS_TIGON_I(ap))
+ writel(link, ®s->TuneFastLink);
+ wmb();
+
+ cmd.evt = C_LNK_NEGOTIATION;
+ cmd.code = 0;
+ cmd.idx = 0;
+ ace_issue_cmd(regs, &cmd);
+ }
+ return 0;
+ }
+#endif
+
+ return -EOPNOTSUPP;
+}
+
+
/*
* Set the hardware MAC address.
*/
-static int ace_set_mac_addr(struct device *dev, void *p)
+static int ace_set_mac_addr(struct net_device *dev, void *p)
{
struct sockaddr *addr=p;
struct ace_regs *regs;
u16 *da;
struct cmd cmd;
- if(dev->start)
+ if(netif_running(dev))
return -EBUSY;
+
memcpy(dev->dev_addr, addr->sa_data,dev->addr_len);
da = (u16 *)dev->dev_addr;
@@ -1548,7 +2555,7 @@
}
-static void ace_set_multicast_list(struct device *dev)
+static void ace_set_multicast_list(struct net_device *dev)
{
struct ace_private *ap = dev->priv;
struct ace_regs *regs = ap->regs;
@@ -1560,9 +2567,9 @@
cmd.idx = 0;
ace_issue_cmd(regs, &cmd);
ap->mcast_all = 1;
- } else if (ap->mcast_all){
+ } else if (ap->mcast_all) {
cmd.evt = C_SET_MULTICAST_MODE;
- cmd.code = C_C_MCAST_ENABLE;
+ cmd.code = C_C_MCAST_DISABLE;
cmd.idx = 0;
ace_issue_cmd(regs, &cmd);
ap->mcast_all = 0;
@@ -1574,7 +2581,7 @@
cmd.idx = 0;
ace_issue_cmd(regs, &cmd);
ap->promisc = 1;
- }else if (!(dev->flags & IFF_PROMISC) && (ap->promisc)){
+ }else if (!(dev->flags & IFF_PROMISC) && (ap->promisc)) {
cmd.evt = C_SET_PROMISC_MODE;
cmd.code = C_C_PROMISC_DISABLE;
cmd.idx = 0;
@@ -1602,15 +2609,21 @@
}
-static struct net_device_stats *ace_get_stats(struct device *dev)
+static struct net_device_stats *ace_get_stats(struct net_device *dev)
{
struct ace_private *ap = dev->priv;
+ struct ace_mac_stats *mac_stats =
+ (struct ace_mac_stats *)ap->regs->Stats;
+
+ ap->stats.rx_missed_errors = readl(&mac_stats->drop_space);
+ ap->stats.multicast = readl(&mac_stats->kept_mc);
+ ap->stats.collisions = readl(&mac_stats->coll);
return(&ap->stats);
}
-__initfunc(void ace_copy(struct ace_regs *regs, void *src, u32 dest, int size))
+void __init ace_copy(struct ace_regs *regs, void *src, u32 dest, int size)
{
unsigned long tdest;
u32 *wsrc;
@@ -1619,23 +2632,20 @@
if (size <= 0)
return;
- while (size > 0){
+ while (size > 0) {
tsize = min(((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
min(size, ACE_WINDOW_SIZE));
tdest = (unsigned long)®s->Window +
(dest & (ACE_WINDOW_SIZE - 1));
writel(dest & ~(ACE_WINDOW_SIZE - 1), ®s->WinBase);
-#ifdef __BIG_ENDIAN
-#error "data must be swapped here"
-#else
-/*
- * XXX - special memcpy needed here!!!
- */
+ /*
+ * This requires byte swapping on big endian, however
+ * writel does that for us
+ */
wsrc = src;
- for (i = 0; i < (tsize / 4); i++){
+ for (i = 0; i < (tsize / 4); i++) {
writel(wsrc[i], tdest + i*4);
}
-#endif
dest += tsize;
src += tsize;
size -= tsize;
@@ -1645,7 +2655,7 @@
}
-__initfunc(void ace_clear(struct ace_regs *regs, u32 dest, int size))
+void __init ace_clear(struct ace_regs *regs, u32 dest, int size)
{
unsigned long tdest;
short tsize = 0, i;
@@ -1653,14 +2663,14 @@
if (size <= 0)
return;
- while (size > 0){
+ while (size > 0) {
tsize = min(((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
min(size, ACE_WINDOW_SIZE));
tdest = (unsigned long)®s->Window +
(dest & (ACE_WINDOW_SIZE - 1));
writel(dest & ~(ACE_WINDOW_SIZE - 1), ®s->WinBase);
- for (i = 0; i < (tsize / 4); i++){
+ for (i = 0; i < (tsize / 4); i++) {
writel(0, tdest + i*4);
}
@@ -1678,15 +2688,15 @@
* This operation requires the NIC to be halted and is performed with
* interrupts disabled and with the spinlock hold.
*/
-__initfunc(int ace_load_firmware(struct device *dev))
+int __init ace_load_firmware(struct net_device *dev)
{
struct ace_private *ap;
struct ace_regs *regs;
- ap = (struct ace_private *)dev->priv;
+ ap = dev->priv;
regs = ap->regs;
- if (!(readl(®s->CpuCtrl) & CPU_HALTED)){
+ if (!(readl(®s->CpuCtrl) & CPU_HALTED)) {
printk(KERN_ERR "%s: trying to download firmware while the "
"CPU is running!\n", dev->name);
return -EFAULT;
@@ -1697,14 +2707,14 @@
* funny things on NICs with only 512KB SRAM
*/
ace_clear(regs, 0x2000, 0x80000-0x2000);
- if (ap->version == 1){
+ if (ACE_IS_TIGON_I(ap)) {
ace_copy(regs, tigonFwText, tigonFwTextAddr, tigonFwTextLen);
ace_copy(regs, tigonFwData, tigonFwDataAddr, tigonFwDataLen);
ace_copy(regs, tigonFwRodata, tigonFwRodataAddr,
tigonFwRodataLen);
ace_clear(regs, tigonFwBssAddr, tigonFwBssLen);
ace_clear(regs, tigonFwSbssAddr, tigonFwSbssLen);
- }else if (ap->version == 2){
+ }else if (ap->version == 2) {
ace_clear(regs, tigon2FwBssAddr, tigon2FwBssLen);
ace_clear(regs, tigon2FwSbssAddr, tigon2FwSbssLen);
ace_copy(regs, tigon2FwText, tigon2FwTextAddr,tigon2FwTextLen);
@@ -1728,36 +2738,40 @@
* specs.
*
* Oh yes, this is only the beginning!
+ *
+ * Thanks to Stevarino Webinski for helping tracking down the bugs in the
+ * code i2c readout code by beta testing all my hacks.
*/
-static void eeprom_start(struct ace_regs *regs)
+static void __init eeprom_start(struct ace_regs *regs)
{
- u32 local = readl(®s->LocalCtrl);
+ u32 local;
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
+ local = readl(®s->LocalCtrl);
local |= EEPROM_DATA_OUT | EEPROM_WRITE_ENABLE;
writel(local, ®s->LocalCtrl);
mb();
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
local |= EEPROM_CLK_OUT;
writel(local, ®s->LocalCtrl);
mb();
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
local &= ~EEPROM_DATA_OUT;
writel(local, ®s->LocalCtrl);
mb();
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
local &= ~EEPROM_CLK_OUT;
writel(local, ®s->LocalCtrl);
mb();
}
-static void eeprom_prep(struct ace_regs *regs, u8 magic)
+static void __init eeprom_prep(struct ace_regs *regs, u8 magic)
{
short i;
u32 local;
- udelay(2);
+ udelay(ACE_SHORT_DELAY);
local = readl(®s->LocalCtrl);
local &= ~EEPROM_DATA_OUT;
local |= EEPROM_WRITE_ENABLE;
@@ -1765,7 +2779,7 @@
mb();
for (i = 0; i < 8; i++, magic <<= 1) {
- udelay(2);
+ udelay(ACE_SHORT_DELAY);
if (magic & 0x80)
local |= EEPROM_DATA_OUT;
else
@@ -1773,11 +2787,11 @@
writel(local, ®s->LocalCtrl);
mb();
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
local |= EEPROM_CLK_OUT;
writel(local, ®s->LocalCtrl);
mb();
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
local &= ~(EEPROM_CLK_OUT | EEPROM_DATA_OUT);
writel(local, ®s->LocalCtrl);
mb();
@@ -1785,7 +2799,7 @@
}
-static int eeprom_check_ack(struct ace_regs *regs)
+static int __init eeprom_check_ack(struct ace_regs *regs)
{
int state;
u32 local;
@@ -1794,14 +2808,14 @@
local &= ~EEPROM_WRITE_ENABLE;
writel(local, ®s->LocalCtrl);
mb();
- udelay(2);
+ udelay(ACE_LONG_DELAY);
local |= EEPROM_CLK_OUT;
writel(local, ®s->LocalCtrl);
mb();
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
/* sample data in middle of high clk */
state = (readl(®s->LocalCtrl) & EEPROM_DATA_IN) != 0;
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
mb();
writel(readl(®s->LocalCtrl) & ~EEPROM_CLK_OUT, ®s->LocalCtrl);
mb();
@@ -1810,27 +2824,28 @@
}
-static void eeprom_stop(struct ace_regs *regs)
+static void __init eeprom_stop(struct ace_regs *regs)
{
u32 local;
+ udelay(ACE_SHORT_DELAY);
local = readl(®s->LocalCtrl);
local |= EEPROM_WRITE_ENABLE;
writel(local, ®s->LocalCtrl);
mb();
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
local &= ~EEPROM_DATA_OUT;
writel(local, ®s->LocalCtrl);
mb();
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
local |= EEPROM_CLK_OUT;
writel(local, ®s->LocalCtrl);
mb();
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
local |= EEPROM_DATA_OUT;
writel(local, ®s->LocalCtrl);
mb();
- udelay(2);
+ udelay(ACE_LONG_DELAY);
local &= ~EEPROM_CLK_OUT;
writel(local, ®s->LocalCtrl);
mb();
@@ -1840,78 +2855,120 @@
/*
* Read a whole byte from the EEPROM.
*/
-static u8 read_eeprom_byte(struct ace_regs *regs, unsigned long offset)
+static int __init read_eeprom_byte(struct net_device *dev,
+ unsigned long offset)
{
+ struct ace_regs *regs;
+ unsigned long flags;
u32 local;
+ int result = 0;
short i;
- u8 result = 0;
- if (!regs){
- printk(KERN_ERR "No regs!\n");
- return 0;
+ if (!dev) {
+ printk(KERN_ERR "No device!\n");
+ result = -ENODEV;
+ goto eeprom_read_error;
}
+ regs = ((struct ace_private *)dev->priv)->regs;
+
+ /*
+ * Don't take interrupts on this CPU will bit banging
+ * the %#%#@$ I2C device
+ */
+ __save_flags(flags);
+ __cli();
+
eeprom_start(regs);
eeprom_prep(regs, EEPROM_WRITE_SELECT);
- if (eeprom_check_ack(regs)){
- printk("Unable to sync eeprom\n");
- return 0;
+ if (eeprom_check_ack(regs)) {
+ __restore_flags(flags);
+ printk(KERN_ERR "%s: Unable to sync eeprom\n", dev->name);
+ result = -EIO;
+ goto eeprom_read_error;
}
eeprom_prep(regs, (offset >> 8) & 0xff);
- if (eeprom_check_ack(regs))
- return 0;
+ if (eeprom_check_ack(regs)) {
+ __restore_flags(flags);
+ printk(KERN_ERR "%s: Unable to set address byte 0\n",
+ dev->name);
+ result = -EIO;
+ goto eeprom_read_error;
+ }
eeprom_prep(regs, offset & 0xff);
- if (eeprom_check_ack(regs))
- return 0;
+ if (eeprom_check_ack(regs)) {
+ __restore_flags(flags);
+ printk(KERN_ERR "%s: Unable to set address byte 1\n",
+ dev->name);
+ result = -EIO;
+ goto eeprom_read_error;
+ }
eeprom_start(regs);
eeprom_prep(regs, EEPROM_READ_SELECT);
- if (eeprom_check_ack(regs))
- return 0;
+ if (eeprom_check_ack(regs)) {
+ __restore_flags(flags);
+ printk(KERN_ERR "%s: Unable to set READ_SELECT\n",
+ dev->name);
+ result = -EIO;
+ goto eeprom_read_error;
+ }
for (i = 0; i < 8; i++) {
local = readl(®s->LocalCtrl);
local &= ~EEPROM_WRITE_ENABLE;
writel(local, ®s->LocalCtrl);
- udelay(2);
+ udelay(ACE_LONG_DELAY);
mb();
local |= EEPROM_CLK_OUT;
writel(local, ®s->LocalCtrl);
- udelay(1);
mb();
+ udelay(ACE_SHORT_DELAY);
/* sample data mid high clk */
result = (result << 1) |
((readl(®s->LocalCtrl) & EEPROM_DATA_IN) != 0);
- udelay(1);
+ udelay(ACE_SHORT_DELAY);
mb();
local = readl(®s->LocalCtrl);
local &= ~EEPROM_CLK_OUT;
writel(local, ®s->LocalCtrl);
+ udelay(ACE_SHORT_DELAY);
mb();
- if (i == 7){
+ if (i == 7) {
local |= EEPROM_WRITE_ENABLE;
writel(local, ®s->LocalCtrl);
mb();
+ udelay(ACE_SHORT_DELAY);
}
}
local |= EEPROM_DATA_OUT;
writel(local, ®s->LocalCtrl);
- udelay(1);
+ mb();
+ udelay(ACE_SHORT_DELAY);
writel(readl(®s->LocalCtrl) | EEPROM_CLK_OUT, ®s->LocalCtrl);
- udelay(2);
+ udelay(ACE_LONG_DELAY);
writel(readl(®s->LocalCtrl) & ~EEPROM_CLK_OUT, ®s->LocalCtrl);
+ mb();
+ udelay(ACE_SHORT_DELAY);
eeprom_stop(regs);
+ __restore_flags(flags);
+ out:
return result;
+
+ eeprom_read_error:
+ printk(KERN_ERR "%s: Unable to read eeprom byte 0x%02lx\n",
+ dev->name, offset);
+ goto out;
}
/*
* Local variables:
- * compile-command: "gcc -D__KERNEL__ -D__SMP__ -DMODULE -I/data/home/jes/linux/include -Wall -Wstrict-prototypes -O2 -fomit-frame-pointer -pipe -fno-strength-reduce -DMODVERSIONS -include /data/home/jes/linux/include/linux/modversions.h -c -o acenic.o acenic.c"
+ * compile-command: "gcc -D__SMP__ -D__KERNEL__ -DMODULE -I../../include -Wall -Wstrict-prototypes -O2 -fomit-frame-pointer -pipe -fno-strength-reduce -DMODVERSIONS -include ../../include/linux/modversions.h -c -o acenic.o acenic.c"
* End:
*/
FUNET's LINUX-ADM group, linux-adm@nic.funet.fi
TCL-scripts by Sam Shen (who was at: slshen@lbl.gov)