2 * acenic.c: Linux driver for the Alteon AceNIC Gigabit Ethernet card
3 * and other Tigon based cards.
5 * Copyright 1998-2002 by Jes Sorensen, <jes@trained-monkey.org>.
7 * Thanks to Alteon and 3Com for providing hardware and documentation
8 * enabling me to write this driver.
10 * A mailing list for discussing the use of this driver has been
11 * setup, please subscribe to the lists if you have any questions
12 * about the driver. Send mail to linux-acenic-help@sunsite.auc.dk to
13 * see how to subscribe.
15 * This program is free software; you can redistribute it and/or modify
16 * it under the terms of the GNU General Public License as published by
17 * the Free Software Foundation; either version 2 of the License, or
18 * (at your option) any later version.
21 * Pete Wyckoff <wyckoff@ca.sandia.gov>: Initial Linux/Alpha and trace
22 * dump support. The trace dump support has not been
23 * integrated yet however.
24 * Troy Benjegerdes: Big Endian (PPC) patches.
25 * Nate Stahl: Better out of memory handling and stats support.
26 * Aman Singla: Nasty race between interrupt handler and tx code dealing
27 * with 'testing the tx_ret_csm and setting tx_full'
28 * David S. Miller <davem@redhat.com>: conversion to new PCI dma mapping
29 * infrastructure and Sparc support
30 * Pierrick Pinasseau (CERN): For lending me an Ultra 5 to test the
31 * driver under Linux/Sparc64
32 * Matt Domsch <Matt_Domsch@dell.com>: Detect Alteon 1000baseT cards
33 * ETHTOOL_GDRVINFO support
34 * Chip Salzenberg <chip@valinux.com>: Fix race condition between tx
35 * handler and close() cleanup.
36 * Ken Aaker <kdaaker@rchland.vnet.ibm.com>: Correct check for whether
37 * memory mapped IO is enabled to
38 * make the driver work on RS/6000.
39 * Takayoshi Kouchi <kouchi@hpc.bs1.fc.nec.co.jp>: Identifying problem
40 * where the driver would disable
41 * bus master mode if it had to disable
42 * write and invalidate.
43 * Stephen Hack <stephen_hack@hp.com>: Fixed ace_set_mac_addr for little
45 * Val Henson <vhenson@esscom.com>: Reset Jumbo skb producer and
46 * rx producer index when
47 * flushing the Jumbo ring.
48 * Hans Grobler <grobh@sun.ac.za>: Memory leak fixes in the
50 * Grant Grundler <grundler@cup.hp.com>: PCI write posting fixes.
53 #include <linux/module.h>
54 #include <linux/moduleparam.h>
55 #include <linux/types.h>
56 #include <linux/errno.h>
57 #include <linux/ioport.h>
58 #include <linux/pci.h>
59 #include <linux/dma-mapping.h>
60 #include <linux/kernel.h>
61 #include <linux/netdevice.h>
62 #include <linux/etherdevice.h>
63 #include <linux/skbuff.h>
64 #include <linux/init.h>
65 #include <linux/delay.h>
67 #include <linux/highmem.h>
68 #include <linux/sockios.h>
70 #if defined(CONFIG_VLAN_8021Q) || defined(CONFIG_VLAN_8021Q_MODULE)
71 #include <linux/if_vlan.h>
75 #include <linux/ethtool.h>
81 #include <asm/system.h>
84 #include <asm/byteorder.h>
85 #include <asm/uaccess.h>
88 #define DRV_NAME "acenic"
92 #ifdef CONFIG_ACENIC_OMIT_TIGON_I
93 #define ACE_IS_TIGON_I(ap) 0
94 #define ACE_TX_RING_ENTRIES(ap) MAX_TX_RING_ENTRIES
96 #define ACE_IS_TIGON_I(ap) (ap->version == 1)
97 #define ACE_TX_RING_ENTRIES(ap) ap->tx_ring_entries
100 #ifndef PCI_VENDOR_ID_ALTEON
101 #define PCI_VENDOR_ID_ALTEON 0x12ae
103 #ifndef PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE
104 #define PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE 0x0001
105 #define PCI_DEVICE_ID_ALTEON_ACENIC_COPPER 0x0002
107 #ifndef PCI_DEVICE_ID_3COM_3C985
108 #define PCI_DEVICE_ID_3COM_3C985 0x0001
110 #ifndef PCI_VENDOR_ID_NETGEAR
111 #define PCI_VENDOR_ID_NETGEAR 0x1385
112 #define PCI_DEVICE_ID_NETGEAR_GA620 0x620a
114 #ifndef PCI_DEVICE_ID_NETGEAR_GA620T
115 #define PCI_DEVICE_ID_NETGEAR_GA620T 0x630a
120 * Farallon used the DEC vendor ID by mistake and they seem not
123 #ifndef PCI_DEVICE_ID_FARALLON_PN9000SX
124 #define PCI_DEVICE_ID_FARALLON_PN9000SX 0x1a
126 #ifndef PCI_DEVICE_ID_FARALLON_PN9100T
127 #define PCI_DEVICE_ID_FARALLON_PN9100T 0xfa
129 #ifndef PCI_VENDOR_ID_SGI
130 #define PCI_VENDOR_ID_SGI 0x10a9
132 #ifndef PCI_DEVICE_ID_SGI_ACENIC
133 #define PCI_DEVICE_ID_SGI_ACENIC 0x0009
136 static struct pci_device_id acenic_pci_tbl[] = {
137 { PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE,
138 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
139 { PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_ALTEON_ACENIC_COPPER,
140 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
141 { PCI_VENDOR_ID_3COM, PCI_DEVICE_ID_3COM_3C985,
142 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
143 { PCI_VENDOR_ID_NETGEAR, PCI_DEVICE_ID_NETGEAR_GA620,
144 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
145 { PCI_VENDOR_ID_NETGEAR, PCI_DEVICE_ID_NETGEAR_GA620T,
146 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
148 * Farallon used the DEC vendor ID on their cards incorrectly,
149 * then later Alteon's ID.
151 { PCI_VENDOR_ID_DEC, PCI_DEVICE_ID_FARALLON_PN9000SX,
152 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
153 { PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_FARALLON_PN9100T,
154 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
155 { PCI_VENDOR_ID_SGI, PCI_DEVICE_ID_SGI_ACENIC,
156 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
159 MODULE_DEVICE_TABLE(pci, acenic_pci_tbl);
161 #define ace_sync_irq(irq) synchronize_irq(irq)
163 #ifndef offset_in_page
164 #define offset_in_page(ptr) ((unsigned long)(ptr) & ~PAGE_MASK)
167 #define ACE_MAX_MOD_PARMS 8
168 #define BOARD_IDX_STATIC 0
169 #define BOARD_IDX_OVERFLOW -1
171 #if (defined(CONFIG_VLAN_8021Q) || defined(CONFIG_VLAN_8021Q_MODULE)) && \
172 defined(NETIF_F_HW_VLAN_RX)
173 #define ACENIC_DO_VLAN 1
174 #define ACE_RCB_VLAN_FLAG RCB_FLG_VLAN_ASSIST
176 #define ACENIC_DO_VLAN 0
177 #define ACE_RCB_VLAN_FLAG 0
183 * These must be defined before the firmware is included.
185 #define MAX_TEXT_LEN 96*1024
186 #define MAX_RODATA_LEN 8*1024
187 #define MAX_DATA_LEN 2*1024
189 #include "acenic_firmware.h"
191 #ifndef tigon2FwReleaseLocal
192 #define tigon2FwReleaseLocal 0
196 * This driver currently supports Tigon I and Tigon II based cards
197 * including the Alteon AceNIC, the 3Com 3C985[B] and NetGear
198 * GA620. The driver should also work on the SGI, DEC and Farallon
199 * versions of the card, however I have not been able to test that
202 * This card is really neat, it supports receive hardware checksumming
203 * and jumbo frames (up to 9000 bytes) and does a lot of work in the
204 * firmware. Also the programming interface is quite neat, except for
205 * the parts dealing with the i2c eeprom on the card ;-)
207 * Using jumbo frames:
209 * To enable jumbo frames, simply specify an mtu between 1500 and 9000
210 * bytes to ifconfig. Jumbo frames can be enabled or disabled at any time
211 * by running `ifconfig eth<X> mtu <MTU>' with <X> being the Ethernet
212 * interface number and <MTU> being the MTU value.
216 * When compiled as a loadable module, the driver allows for a number
217 * of module parameters to be specified. The driver supports the
218 * following module parameters:
220 * trace=<val> - Firmware trace level. This requires special traced
221 * firmware to replace the firmware supplied with
222 * the driver - for debugging purposes only.
224 * link=<val> - Link state. Normally you want to use the default link
225 * parameters set by the driver. This can be used to
226 * override these in case your switch doesn't negotiate
227 * the link properly. Valid values are:
228 * 0x0001 - Force half duplex link.
229 * 0x0002 - Do not negotiate line speed with the other end.
230 * 0x0010 - 10Mbit/sec link.
231 * 0x0020 - 100Mbit/sec link.
232 * 0x0040 - 1000Mbit/sec link.
233 * 0x0100 - Do not negotiate flow control.
234 * 0x0200 - Enable RX flow control Y
235 * 0x0400 - Enable TX flow control Y (Tigon II NICs only).
236 * Default value is 0x0270, ie. enable link+flow
237 * control negotiation. Negotiating the highest
238 * possible link speed with RX flow control enabled.
240 * When disabling link speed negotiation, only one link
241 * speed is allowed to be specified!
243 * tx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
244 * to wait for more packets to arive before
245 * interrupting the host, from the time the first
248 * rx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
249 * to wait for more packets to arive in the transmit ring,
250 * before interrupting the host, after transmitting the
251 * first packet in the ring.
253 * max_tx_desc=<val> - maximum number of transmit descriptors
254 * (packets) transmitted before interrupting the host.
256 * max_rx_desc=<val> - maximum number of receive descriptors
257 * (packets) received before interrupting the host.
259 * tx_ratio=<val> - 7 bit value (0 - 63) specifying the split in 64th
260 * increments of the NIC's on board memory to be used for
261 * transmit and receive buffers. For the 1MB NIC app. 800KB
262 * is available, on the 1/2MB NIC app. 300KB is available.
263 * 68KB will always be available as a minimum for both
264 * directions. The default value is a 50/50 split.
265 * dis_pci_mem_inval=<val> - disable PCI memory write and invalidate
266 * operations, default (1) is to always disable this as
267 * that is what Alteon does on NT. I have not been able
268 * to measure any real performance differences with
269 * this on my systems. Set <val>=0 if you want to
270 * enable these operations.
272 * If you use more than one NIC, specify the parameters for the
273 * individual NICs with a comma, ie. trace=0,0x00001fff,0 you want to
274 * run tracing on NIC #2 but not on NIC #1 and #3.
278 * - Proper multicast support.
279 * - NIC dump support.
280 * - More tuning parameters.
282 * The mini ring is not used under Linux and I am not sure it makes sense
283 * to actually use it.
285 * New interrupt handler strategy:
287 * The old interrupt handler worked using the traditional method of
288 * replacing an skbuff with a new one when a packet arrives. However
289 * the rx rings do not need to contain a static number of buffer
290 * descriptors, thus it makes sense to move the memory allocation out
291 * of the main interrupt handler and do it in a bottom half handler
292 * and only allocate new buffers when the number of buffers in the
293 * ring is below a certain threshold. In order to avoid starving the
294 * NIC under heavy load it is however necessary to force allocation
295 * when hitting a minimum threshold. The strategy for alloction is as
298 * RX_LOW_BUF_THRES - allocate buffers in the bottom half
299 * RX_PANIC_LOW_THRES - we are very low on buffers, allocate
300 * the buffers in the interrupt handler
301 * RX_RING_THRES - maximum number of buffers in the rx ring
302 * RX_MINI_THRES - maximum number of buffers in the mini ring
303 * RX_JUMBO_THRES - maximum number of buffers in the jumbo ring
305 * One advantagous side effect of this allocation approach is that the
306 * entire rx processing can be done without holding any spin lock
307 * since the rx rings and registers are totally independent of the tx
308 * ring and its registers. This of course includes the kmalloc's of
309 * new skb's. Thus start_xmit can run in parallel with rx processing
310 * and the memory allocation on SMP systems.
312 * Note that running the skb reallocation in a bottom half opens up
313 * another can of races which needs to be handled properly. In
314 * particular it can happen that the interrupt handler tries to run
315 * the reallocation while the bottom half is either running on another
316 * CPU or was interrupted on the same CPU. To get around this the
317 * driver uses bitops to prevent the reallocation routines from being
320 * TX handling can also be done without holding any spin lock, wheee
321 * this is fun! since tx_ret_csm is only written to by the interrupt
322 * handler. The case to be aware of is when shutting down the device
323 * and cleaning up where it is necessary to make sure that
324 * start_xmit() is not running while this is happening. Well DaveM
325 * informs me that this case is already protected against ... bye bye
326 * Mr. Spin Lock, it was nice to know you.
328 * TX interrupts are now partly disabled so the NIC will only generate
329 * TX interrupts for the number of coal ticks, not for the number of
330 * TX packets in the queue. This should reduce the number of TX only,
331 * ie. when no RX processing is done, interrupts seen.
335 * Threshold values for RX buffer allocation - the low water marks for
336 * when to start refilling the rings are set to 75% of the ring
337 * sizes. It seems to make sense to refill the rings entirely from the
338 * intrrupt handler once it gets below the panic threshold, that way
339 * we don't risk that the refilling is moved to another CPU when the
340 * one running the interrupt handler just got the slab code hot in its
343 #define RX_RING_SIZE 72
344 #define RX_MINI_SIZE 64
345 #define RX_JUMBO_SIZE 48
347 #define RX_PANIC_STD_THRES 16
348 #define RX_PANIC_STD_REFILL (3*RX_PANIC_STD_THRES)/2
349 #define RX_LOW_STD_THRES (3*RX_RING_SIZE)/4
350 #define RX_PANIC_MINI_THRES 12
351 #define RX_PANIC_MINI_REFILL (3*RX_PANIC_MINI_THRES)/2
352 #define RX_LOW_MINI_THRES (3*RX_MINI_SIZE)/4
353 #define RX_PANIC_JUMBO_THRES 6
354 #define RX_PANIC_JUMBO_REFILL (3*RX_PANIC_JUMBO_THRES)/2
355 #define RX_LOW_JUMBO_THRES (3*RX_JUMBO_SIZE)/4
359 * Size of the mini ring entries, basically these just should be big
360 * enough to take TCP ACKs
362 #define ACE_MINI_SIZE 100
364 #define ACE_MINI_BUFSIZE ACE_MINI_SIZE
365 #define ACE_STD_BUFSIZE (ACE_STD_MTU + ETH_HLEN + 4)
366 #define ACE_JUMBO_BUFSIZE (ACE_JUMBO_MTU + ETH_HLEN + 4)
369 * There seems to be a magic difference in the effect between 995 and 996
370 * but little difference between 900 and 995 ... no idea why.
372 * There is now a default set of tuning parameters which is set, depending
373 * on whether or not the user enables Jumbo frames. It's assumed that if
374 * Jumbo frames are enabled, the user wants optimal tuning for that case.
376 #define DEF_TX_COAL 400 /* 996 */
377 #define DEF_TX_MAX_DESC 60 /* was 40 */
378 #define DEF_RX_COAL 120 /* 1000 */
379 #define DEF_RX_MAX_DESC 25
380 #define DEF_TX_RATIO 21 /* 24 */
382 #define DEF_JUMBO_TX_COAL 20
383 #define DEF_JUMBO_TX_MAX_DESC 60
384 #define DEF_JUMBO_RX_COAL 30
385 #define DEF_JUMBO_RX_MAX_DESC 6
386 #define DEF_JUMBO_TX_RATIO 21
388 #if tigon2FwReleaseLocal < 20001118
390 * Standard firmware and early modifications duplicate
391 * IRQ load without this flag (coal timer is never reset).
392 * Note that with this flag tx_coal should be less than
393 * time to xmit full tx ring.
394 * 400usec is not so bad for tx ring size of 128.
396 #define TX_COAL_INTS_ONLY 1 /* worth it */
399 * With modified firmware, this is not necessary, but still useful.
401 #define TX_COAL_INTS_ONLY 1
405 #define DEF_STAT (2 * TICKS_PER_SEC)
408 static int link_state[ACE_MAX_MOD_PARMS];
409 static int trace[ACE_MAX_MOD_PARMS];
410 static int tx_coal_tick[ACE_MAX_MOD_PARMS];
411 static int rx_coal_tick[ACE_MAX_MOD_PARMS];
412 static int max_tx_desc[ACE_MAX_MOD_PARMS];
413 static int max_rx_desc[ACE_MAX_MOD_PARMS];
414 static int tx_ratio[ACE_MAX_MOD_PARMS];
415 static int dis_pci_mem_inval[ACE_MAX_MOD_PARMS] = {1, 1, 1, 1, 1, 1, 1, 1};
417 MODULE_AUTHOR("Jes Sorensen <jes@trained-monkey.org>");
418 MODULE_LICENSE("GPL");
419 MODULE_DESCRIPTION("AceNIC/3C985/GA620 Gigabit Ethernet driver");
421 module_param_array_named(link, link_state, int, NULL, 0);
422 module_param_array(trace, int, NULL, 0);
423 module_param_array(tx_coal_tick, int, NULL, 0);
424 module_param_array(max_tx_desc, int, NULL, 0);
425 module_param_array(rx_coal_tick, int, NULL, 0);
426 module_param_array(max_rx_desc, int, NULL, 0);
427 module_param_array(tx_ratio, int, NULL, 0);
428 MODULE_PARM_DESC(link, "AceNIC/3C985/NetGear link state");
429 MODULE_PARM_DESC(trace, "AceNIC/3C985/NetGear firmware trace level");
430 MODULE_PARM_DESC(tx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first tx descriptor arrives");
431 MODULE_PARM_DESC(max_tx_desc, "AceNIC/3C985/GA620 max number of transmit descriptors to wait");
432 MODULE_PARM_DESC(rx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first rx descriptor arrives");
433 MODULE_PARM_DESC(max_rx_desc, "AceNIC/3C985/GA620 max number of receive descriptors to wait");
434 MODULE_PARM_DESC(tx_ratio, "AceNIC/3C985/GA620 ratio of NIC memory used for TX/RX descriptors (range 0-63)");
437 static char version[] __devinitdata =
438 "acenic.c: v0.92 08/05/2002 Jes Sorensen, linux-acenic@SunSITE.dk\n"
439 " http://home.cern.ch/~jes/gige/acenic.html\n";
441 static int ace_get_settings(struct net_device *, struct ethtool_cmd *);
442 static int ace_set_settings(struct net_device *, struct ethtool_cmd *);
443 static void ace_get_drvinfo(struct net_device *, struct ethtool_drvinfo *);
445 static const struct ethtool_ops ace_ethtool_ops = {
446 .get_settings = ace_get_settings,
447 .set_settings = ace_set_settings,
448 .get_drvinfo = ace_get_drvinfo,
451 static void ace_watchdog(struct net_device *dev);
453 static const struct net_device_ops ace_netdev_ops = {
454 .ndo_open = ace_open,
455 .ndo_stop = ace_close,
456 .ndo_tx_timeout = ace_watchdog,
457 .ndo_get_stats = ace_get_stats,
458 .ndo_start_xmit = ace_start_xmit,
459 .ndo_set_multicast_list = ace_set_multicast_list,
460 .ndo_set_mac_address = ace_set_mac_addr,
461 .ndo_change_mtu = ace_change_mtu,
463 .ndo_vlan_rx_register = ace_vlan_rx_register,
467 static int __devinit acenic_probe_one(struct pci_dev *pdev,
468 const struct pci_device_id *id)
470 struct net_device *dev;
471 struct ace_private *ap;
472 static int boards_found;
474 dev = alloc_etherdev(sizeof(struct ace_private));
476 printk(KERN_ERR "acenic: Unable to allocate "
477 "net_device structure!\n");
481 SET_NETDEV_DEV(dev, &pdev->dev);
483 ap = netdev_priv(dev);
485 ap->name = pci_name(pdev);
487 dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
489 dev->features |= NETIF_F_HW_VLAN_TX | NETIF_F_HW_VLAN_RX;
492 dev->watchdog_timeo = 5*HZ;
494 dev->netdev_ops = &ace_netdev_ops;
495 SET_ETHTOOL_OPS(dev, &ace_ethtool_ops);
497 /* we only display this string ONCE */
501 if (pci_enable_device(pdev))
502 goto fail_free_netdev;
505 * Enable master mode before we start playing with the
506 * pci_command word since pci_set_master() will modify
509 pci_set_master(pdev);
511 pci_read_config_word(pdev, PCI_COMMAND, &ap->pci_command);
513 /* OpenFirmware on Mac's does not set this - DOH.. */
514 if (!(ap->pci_command & PCI_COMMAND_MEMORY)) {
515 printk(KERN_INFO "%s: Enabling PCI Memory Mapped "
516 "access - was not enabled by BIOS/Firmware\n",
518 ap->pci_command = ap->pci_command | PCI_COMMAND_MEMORY;
519 pci_write_config_word(ap->pdev, PCI_COMMAND,
524 pci_read_config_byte(pdev, PCI_LATENCY_TIMER, &ap->pci_latency);
525 if (ap->pci_latency <= 0x40) {
526 ap->pci_latency = 0x40;
527 pci_write_config_byte(pdev, PCI_LATENCY_TIMER, ap->pci_latency);
531 * Remap the regs into kernel space - this is abuse of
532 * dev->base_addr since it was means for I/O port
533 * addresses but who gives a damn.
535 dev->base_addr = pci_resource_start(pdev, 0);
536 ap->regs = ioremap(dev->base_addr, 0x4000);
538 printk(KERN_ERR "%s: Unable to map I/O register, "
539 "AceNIC %i will be disabled.\n",
540 ap->name, boards_found);
541 goto fail_free_netdev;
544 switch(pdev->vendor) {
545 case PCI_VENDOR_ID_ALTEON:
546 if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9100T) {
547 printk(KERN_INFO "%s: Farallon PN9100-T ",
550 printk(KERN_INFO "%s: Alteon AceNIC ",
554 case PCI_VENDOR_ID_3COM:
555 printk(KERN_INFO "%s: 3Com 3C985 ", ap->name);
557 case PCI_VENDOR_ID_NETGEAR:
558 printk(KERN_INFO "%s: NetGear GA620 ", ap->name);
560 case PCI_VENDOR_ID_DEC:
561 if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9000SX) {
562 printk(KERN_INFO "%s: Farallon PN9000-SX ",
566 case PCI_VENDOR_ID_SGI:
567 printk(KERN_INFO "%s: SGI AceNIC ", ap->name);
570 printk(KERN_INFO "%s: Unknown AceNIC ", ap->name);
574 printk("Gigabit Ethernet at 0x%08lx, ", dev->base_addr);
575 printk("irq %d\n", pdev->irq);
577 #ifdef CONFIG_ACENIC_OMIT_TIGON_I
578 if ((readl(&ap->regs->HostCtrl) >> 28) == 4) {
579 printk(KERN_ERR "%s: Driver compiled without Tigon I"
580 " support - NIC disabled\n", dev->name);
585 if (ace_allocate_descriptors(dev))
586 goto fail_free_netdev;
589 if (boards_found >= ACE_MAX_MOD_PARMS)
590 ap->board_idx = BOARD_IDX_OVERFLOW;
592 ap->board_idx = boards_found;
594 ap->board_idx = BOARD_IDX_STATIC;
598 goto fail_free_netdev;
600 if (register_netdev(dev)) {
601 printk(KERN_ERR "acenic: device registration failed\n");
604 ap->name = dev->name;
606 if (ap->pci_using_dac)
607 dev->features |= NETIF_F_HIGHDMA;
609 pci_set_drvdata(pdev, dev);
615 ace_init_cleanup(dev);
621 static void __devexit acenic_remove_one(struct pci_dev *pdev)
623 struct net_device *dev = pci_get_drvdata(pdev);
624 struct ace_private *ap = netdev_priv(dev);
625 struct ace_regs __iomem *regs = ap->regs;
628 unregister_netdev(dev);
630 writel(readl(®s->CpuCtrl) | CPU_HALT, ®s->CpuCtrl);
631 if (ap->version >= 2)
632 writel(readl(®s->CpuBCtrl) | CPU_HALT, ®s->CpuBCtrl);
635 * This clears any pending interrupts
637 writel(1, ®s->Mb0Lo);
638 readl(®s->CpuCtrl); /* flush */
641 * Make sure no other CPUs are processing interrupts
642 * on the card before the buffers are being released.
643 * Otherwise one might experience some `interesting'
646 * Then release the RX buffers - jumbo buffers were
647 * already released in ace_close().
649 ace_sync_irq(dev->irq);
651 for (i = 0; i < RX_STD_RING_ENTRIES; i++) {
652 struct sk_buff *skb = ap->skb->rx_std_skbuff[i].skb;
655 struct ring_info *ringp;
658 ringp = &ap->skb->rx_std_skbuff[i];
659 mapping = pci_unmap_addr(ringp, mapping);
660 pci_unmap_page(ap->pdev, mapping,
664 ap->rx_std_ring[i].size = 0;
665 ap->skb->rx_std_skbuff[i].skb = NULL;
670 if (ap->version >= 2) {
671 for (i = 0; i < RX_MINI_RING_ENTRIES; i++) {
672 struct sk_buff *skb = ap->skb->rx_mini_skbuff[i].skb;
675 struct ring_info *ringp;
678 ringp = &ap->skb->rx_mini_skbuff[i];
679 mapping = pci_unmap_addr(ringp,mapping);
680 pci_unmap_page(ap->pdev, mapping,
684 ap->rx_mini_ring[i].size = 0;
685 ap->skb->rx_mini_skbuff[i].skb = NULL;
691 for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
692 struct sk_buff *skb = ap->skb->rx_jumbo_skbuff[i].skb;
694 struct ring_info *ringp;
697 ringp = &ap->skb->rx_jumbo_skbuff[i];
698 mapping = pci_unmap_addr(ringp, mapping);
699 pci_unmap_page(ap->pdev, mapping,
703 ap->rx_jumbo_ring[i].size = 0;
704 ap->skb->rx_jumbo_skbuff[i].skb = NULL;
709 ace_init_cleanup(dev);
713 static struct pci_driver acenic_pci_driver = {
715 .id_table = acenic_pci_tbl,
716 .probe = acenic_probe_one,
717 .remove = __devexit_p(acenic_remove_one),
720 static int __init acenic_init(void)
722 return pci_register_driver(&acenic_pci_driver);
725 static void __exit acenic_exit(void)
727 pci_unregister_driver(&acenic_pci_driver);
730 module_init(acenic_init);
731 module_exit(acenic_exit);
733 static void ace_free_descriptors(struct net_device *dev)
735 struct ace_private *ap = netdev_priv(dev);
738 if (ap->rx_std_ring != NULL) {
739 size = (sizeof(struct rx_desc) *
740 (RX_STD_RING_ENTRIES +
741 RX_JUMBO_RING_ENTRIES +
742 RX_MINI_RING_ENTRIES +
743 RX_RETURN_RING_ENTRIES));
744 pci_free_consistent(ap->pdev, size, ap->rx_std_ring,
745 ap->rx_ring_base_dma);
746 ap->rx_std_ring = NULL;
747 ap->rx_jumbo_ring = NULL;
748 ap->rx_mini_ring = NULL;
749 ap->rx_return_ring = NULL;
751 if (ap->evt_ring != NULL) {
752 size = (sizeof(struct event) * EVT_RING_ENTRIES);
753 pci_free_consistent(ap->pdev, size, ap->evt_ring,
757 if (ap->tx_ring != NULL && !ACE_IS_TIGON_I(ap)) {
758 size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
759 pci_free_consistent(ap->pdev, size, ap->tx_ring,
764 if (ap->evt_prd != NULL) {
765 pci_free_consistent(ap->pdev, sizeof(u32),
766 (void *)ap->evt_prd, ap->evt_prd_dma);
769 if (ap->rx_ret_prd != NULL) {
770 pci_free_consistent(ap->pdev, sizeof(u32),
771 (void *)ap->rx_ret_prd,
773 ap->rx_ret_prd = NULL;
775 if (ap->tx_csm != NULL) {
776 pci_free_consistent(ap->pdev, sizeof(u32),
777 (void *)ap->tx_csm, ap->tx_csm_dma);
783 static int ace_allocate_descriptors(struct net_device *dev)
785 struct ace_private *ap = netdev_priv(dev);
788 size = (sizeof(struct rx_desc) *
789 (RX_STD_RING_ENTRIES +
790 RX_JUMBO_RING_ENTRIES +
791 RX_MINI_RING_ENTRIES +
792 RX_RETURN_RING_ENTRIES));
794 ap->rx_std_ring = pci_alloc_consistent(ap->pdev, size,
795 &ap->rx_ring_base_dma);
796 if (ap->rx_std_ring == NULL)
799 ap->rx_jumbo_ring = ap->rx_std_ring + RX_STD_RING_ENTRIES;
800 ap->rx_mini_ring = ap->rx_jumbo_ring + RX_JUMBO_RING_ENTRIES;
801 ap->rx_return_ring = ap->rx_mini_ring + RX_MINI_RING_ENTRIES;
803 size = (sizeof(struct event) * EVT_RING_ENTRIES);
805 ap->evt_ring = pci_alloc_consistent(ap->pdev, size, &ap->evt_ring_dma);
807 if (ap->evt_ring == NULL)
811 * Only allocate a host TX ring for the Tigon II, the Tigon I
812 * has to use PCI registers for this ;-(
814 if (!ACE_IS_TIGON_I(ap)) {
815 size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
817 ap->tx_ring = pci_alloc_consistent(ap->pdev, size,
820 if (ap->tx_ring == NULL)
824 ap->evt_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
826 if (ap->evt_prd == NULL)
829 ap->rx_ret_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
830 &ap->rx_ret_prd_dma);
831 if (ap->rx_ret_prd == NULL)
834 ap->tx_csm = pci_alloc_consistent(ap->pdev, sizeof(u32),
836 if (ap->tx_csm == NULL)
843 ace_init_cleanup(dev);
849 * Generic cleanup handling data allocated during init. Used when the
850 * module is unloaded or if an error occurs during initialization
852 static void ace_init_cleanup(struct net_device *dev)
854 struct ace_private *ap;
856 ap = netdev_priv(dev);
858 ace_free_descriptors(dev);
861 pci_free_consistent(ap->pdev, sizeof(struct ace_info),
862 ap->info, ap->info_dma);
864 kfree(ap->trace_buf);
867 free_irq(dev->irq, dev);
874 * Commands are considered to be slow.
876 static inline void ace_issue_cmd(struct ace_regs __iomem *regs, struct cmd *cmd)
880 idx = readl(®s->CmdPrd);
882 writel(*(u32 *)(cmd), ®s->CmdRng[idx]);
883 idx = (idx + 1) % CMD_RING_ENTRIES;
885 writel(idx, ®s->CmdPrd);
889 static int __devinit ace_init(struct net_device *dev)
891 struct ace_private *ap;
892 struct ace_regs __iomem *regs;
893 struct ace_info *info = NULL;
894 struct pci_dev *pdev;
897 u32 tig_ver, mac1, mac2, tmp, pci_state;
898 int board_idx, ecode = 0;
900 unsigned char cache_size;
902 ap = netdev_priv(dev);
905 board_idx = ap->board_idx;
908 * aman@sgi.com - its useful to do a NIC reset here to
909 * address the `Firmware not running' problem subsequent
910 * to any crashes involving the NIC
912 writel(HW_RESET | (HW_RESET << 24), ®s->HostCtrl);
913 readl(®s->HostCtrl); /* PCI write posting */
917 * Don't access any other registers before this point!
921 * This will most likely need BYTE_SWAP once we switch
922 * to using __raw_writel()
924 writel((WORD_SWAP | CLR_INT | ((WORD_SWAP | CLR_INT) << 24)),
927 writel((CLR_INT | WORD_SWAP | ((CLR_INT | WORD_SWAP) << 24)),
930 readl(®s->HostCtrl); /* PCI write posting */
933 * Stop the NIC CPU and clear pending interrupts
935 writel(readl(®s->CpuCtrl) | CPU_HALT, ®s->CpuCtrl);
936 readl(®s->CpuCtrl); /* PCI write posting */
937 writel(0, ®s->Mb0Lo);
939 tig_ver = readl(®s->HostCtrl) >> 28;
942 #ifndef CONFIG_ACENIC_OMIT_TIGON_I
945 printk(KERN_INFO " Tigon I (Rev. %i), Firmware: %i.%i.%i, ",
946 tig_ver, tigonFwReleaseMajor, tigonFwReleaseMinor,
948 writel(0, ®s->LocalCtrl);
950 ap->tx_ring_entries = TIGON_I_TX_RING_ENTRIES;
954 printk(KERN_INFO " Tigon II (Rev. %i), Firmware: %i.%i.%i, ",
955 tig_ver, tigon2FwReleaseMajor, tigon2FwReleaseMinor,
957 writel(readl(®s->CpuBCtrl) | CPU_HALT, ®s->CpuBCtrl);
958 readl(®s->CpuBCtrl); /* PCI write posting */
960 * The SRAM bank size does _not_ indicate the amount
961 * of memory on the card, it controls the _bank_ size!
962 * Ie. a 1MB AceNIC will have two banks of 512KB.
964 writel(SRAM_BANK_512K, ®s->LocalCtrl);
965 writel(SYNC_SRAM_TIMING, ®s->MiscCfg);
967 ap->tx_ring_entries = MAX_TX_RING_ENTRIES;
970 printk(KERN_WARNING " Unsupported Tigon version detected "
977 * ModeStat _must_ be set after the SRAM settings as this change
978 * seems to corrupt the ModeStat and possible other registers.
979 * The SRAM settings survive resets and setting it to the same
980 * value a second time works as well. This is what caused the
981 * `Firmware not running' problem on the Tigon II.
984 writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL | ACE_BYTE_SWAP_BD |
985 ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, ®s->ModeStat);
987 writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL |
988 ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, ®s->ModeStat);
990 readl(®s->ModeStat); /* PCI write posting */
993 for(i = 0; i < 4; i++) {
997 t = read_eeprom_byte(dev, 0x8c+i);
1005 for(i = 4; i < 8; i++) {
1009 t = read_eeprom_byte(dev, 0x8c+i);
1017 writel(mac1, ®s->MacAddrHi);
1018 writel(mac2, ®s->MacAddrLo);
1020 dev->dev_addr[0] = (mac1 >> 8) & 0xff;
1021 dev->dev_addr[1] = mac1 & 0xff;
1022 dev->dev_addr[2] = (mac2 >> 24) & 0xff;
1023 dev->dev_addr[3] = (mac2 >> 16) & 0xff;
1024 dev->dev_addr[4] = (mac2 >> 8) & 0xff;
1025 dev->dev_addr[5] = mac2 & 0xff;
1027 printk("MAC: %pM\n", dev->dev_addr);
1030 * Looks like this is necessary to deal with on all architectures,
1031 * even this %$#%$# N440BX Intel based thing doesn't get it right.
1032 * Ie. having two NICs in the machine, one will have the cache
1033 * line set at boot time, the other will not.
1036 pci_read_config_byte(pdev, PCI_CACHE_LINE_SIZE, &cache_size);
1038 if (cache_size != SMP_CACHE_BYTES) {
1039 printk(KERN_INFO " PCI cache line size set incorrectly "
1040 "(%i bytes) by BIOS/FW, ", cache_size);
1041 if (cache_size > SMP_CACHE_BYTES)
1042 printk("expecting %i\n", SMP_CACHE_BYTES);
1044 printk("correcting to %i\n", SMP_CACHE_BYTES);
1045 pci_write_config_byte(pdev, PCI_CACHE_LINE_SIZE,
1046 SMP_CACHE_BYTES >> 2);
1050 pci_state = readl(®s->PciState);
1051 printk(KERN_INFO " PCI bus width: %i bits, speed: %iMHz, "
1052 "latency: %i clks\n",
1053 (pci_state & PCI_32BIT) ? 32 : 64,
1054 (pci_state & PCI_66MHZ) ? 66 : 33,
1058 * Set the max DMA transfer size. Seems that for most systems
1059 * the performance is better when no MAX parameter is
1060 * set. However for systems enabling PCI write and invalidate,
1061 * DMA writes must be set to the L1 cache line size to get
1062 * optimal performance.
1064 * The default is now to turn the PCI write and invalidate off
1065 * - that is what Alteon does for NT.
1067 tmp = READ_CMD_MEM | WRITE_CMD_MEM;
1068 if (ap->version >= 2) {
1069 tmp |= (MEM_READ_MULTIPLE | (pci_state & PCI_66MHZ));
1071 * Tuning parameters only supported for 8 cards
1073 if (board_idx == BOARD_IDX_OVERFLOW ||
1074 dis_pci_mem_inval[board_idx]) {
1075 if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1076 ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1077 pci_write_config_word(pdev, PCI_COMMAND,
1079 printk(KERN_INFO " Disabling PCI memory "
1080 "write and invalidate\n");
1082 } else if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1083 printk(KERN_INFO " PCI memory write & invalidate "
1084 "enabled by BIOS, enabling counter measures\n");
1086 switch(SMP_CACHE_BYTES) {
1088 tmp |= DMA_WRITE_MAX_16;
1091 tmp |= DMA_WRITE_MAX_32;
1094 tmp |= DMA_WRITE_MAX_64;
1097 tmp |= DMA_WRITE_MAX_128;
1100 printk(KERN_INFO " Cache line size %i not "
1101 "supported, PCI write and invalidate "
1102 "disabled\n", SMP_CACHE_BYTES);
1103 ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1104 pci_write_config_word(pdev, PCI_COMMAND,
1112 * On this platform, we know what the best dma settings
1113 * are. We use 64-byte maximum bursts, because if we
1114 * burst larger than the cache line size (or even cross
1115 * a 64byte boundary in a single burst) the UltraSparc
1116 * PCI controller will disconnect at 64-byte multiples.
1118 * Read-multiple will be properly enabled above, and when
1119 * set will give the PCI controller proper hints about
1122 tmp &= ~DMA_READ_WRITE_MASK;
1123 tmp |= DMA_READ_MAX_64;
1124 tmp |= DMA_WRITE_MAX_64;
1127 tmp &= ~DMA_READ_WRITE_MASK;
1128 tmp |= DMA_READ_MAX_128;
1130 * All the docs say MUST NOT. Well, I did.
1131 * Nothing terrible happens, if we load wrong size.
1132 * Bit w&i still works better!
1134 tmp |= DMA_WRITE_MAX_128;
1136 writel(tmp, ®s->PciState);
1140 * The Host PCI bus controller driver has to set FBB.
1141 * If all devices on that PCI bus support FBB, then the controller
1142 * can enable FBB support in the Host PCI Bus controller (or on
1143 * the PCI-PCI bridge if that applies).
1147 * I have received reports from people having problems when this
1150 if (!(ap->pci_command & PCI_COMMAND_FAST_BACK)) {
1151 printk(KERN_INFO " Enabling PCI Fast Back to Back\n");
1152 ap->pci_command |= PCI_COMMAND_FAST_BACK;
1153 pci_write_config_word(pdev, PCI_COMMAND, ap->pci_command);
1158 * Configure DMA attributes.
1160 if (!pci_set_dma_mask(pdev, DMA_64BIT_MASK)) {
1161 ap->pci_using_dac = 1;
1162 } else if (!pci_set_dma_mask(pdev, DMA_32BIT_MASK)) {
1163 ap->pci_using_dac = 0;
1170 * Initialize the generic info block and the command+event rings
1171 * and the control blocks for the transmit and receive rings
1172 * as they need to be setup once and for all.
1174 if (!(info = pci_alloc_consistent(ap->pdev, sizeof(struct ace_info),
1182 * Get the memory for the skb rings.
1184 if (!(ap->skb = kmalloc(sizeof(struct ace_skb), GFP_KERNEL))) {
1189 ecode = request_irq(pdev->irq, ace_interrupt, IRQF_SHARED,
1192 printk(KERN_WARNING "%s: Requested IRQ %d is busy\n",
1193 DRV_NAME, pdev->irq);
1196 dev->irq = pdev->irq;
1199 spin_lock_init(&ap->debug_lock);
1200 ap->last_tx = ACE_TX_RING_ENTRIES(ap) - 1;
1201 ap->last_std_rx = 0;
1202 ap->last_mini_rx = 0;
1205 memset(ap->info, 0, sizeof(struct ace_info));
1206 memset(ap->skb, 0, sizeof(struct ace_skb));
1208 ace_load_firmware(dev);
1211 tmp_ptr = ap->info_dma;
1212 writel(tmp_ptr >> 32, ®s->InfoPtrHi);
1213 writel(tmp_ptr & 0xffffffff, ®s->InfoPtrLo);
1215 memset(ap->evt_ring, 0, EVT_RING_ENTRIES * sizeof(struct event));
1217 set_aceaddr(&info->evt_ctrl.rngptr, ap->evt_ring_dma);
1218 info->evt_ctrl.flags = 0;
1222 set_aceaddr(&info->evt_prd_ptr, ap->evt_prd_dma);
1223 writel(0, ®s->EvtCsm);
1225 set_aceaddr(&info->cmd_ctrl.rngptr, 0x100);
1226 info->cmd_ctrl.flags = 0;
1227 info->cmd_ctrl.max_len = 0;
1229 for (i = 0; i < CMD_RING_ENTRIES; i++)
1230 writel(0, ®s->CmdRng[i]);
1232 writel(0, ®s->CmdPrd);
1233 writel(0, ®s->CmdCsm);
1235 tmp_ptr = ap->info_dma;
1236 tmp_ptr += (unsigned long) &(((struct ace_info *)0)->s.stats);
1237 set_aceaddr(&info->stats2_ptr, (dma_addr_t) tmp_ptr);
1239 set_aceaddr(&info->rx_std_ctrl.rngptr, ap->rx_ring_base_dma);
1240 info->rx_std_ctrl.max_len = ACE_STD_BUFSIZE;
1241 info->rx_std_ctrl.flags =
1242 RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | ACE_RCB_VLAN_FLAG;
1244 memset(ap->rx_std_ring, 0,
1245 RX_STD_RING_ENTRIES * sizeof(struct rx_desc));
1247 for (i = 0; i < RX_STD_RING_ENTRIES; i++)
1248 ap->rx_std_ring[i].flags = BD_FLG_TCP_UDP_SUM;
1250 ap->rx_std_skbprd = 0;
1251 atomic_set(&ap->cur_rx_bufs, 0);
1253 set_aceaddr(&info->rx_jumbo_ctrl.rngptr,
1254 (ap->rx_ring_base_dma +
1255 (sizeof(struct rx_desc) * RX_STD_RING_ENTRIES)));
1256 info->rx_jumbo_ctrl.max_len = 0;
1257 info->rx_jumbo_ctrl.flags =
1258 RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | ACE_RCB_VLAN_FLAG;
1260 memset(ap->rx_jumbo_ring, 0,
1261 RX_JUMBO_RING_ENTRIES * sizeof(struct rx_desc));
1263 for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++)
1264 ap->rx_jumbo_ring[i].flags = BD_FLG_TCP_UDP_SUM | BD_FLG_JUMBO;
1266 ap->rx_jumbo_skbprd = 0;
1267 atomic_set(&ap->cur_jumbo_bufs, 0);
1269 memset(ap->rx_mini_ring, 0,
1270 RX_MINI_RING_ENTRIES * sizeof(struct rx_desc));
1272 if (ap->version >= 2) {
1273 set_aceaddr(&info->rx_mini_ctrl.rngptr,
1274 (ap->rx_ring_base_dma +
1275 (sizeof(struct rx_desc) *
1276 (RX_STD_RING_ENTRIES +
1277 RX_JUMBO_RING_ENTRIES))));
1278 info->rx_mini_ctrl.max_len = ACE_MINI_SIZE;
1279 info->rx_mini_ctrl.flags =
1280 RCB_FLG_TCP_UDP_SUM|RCB_FLG_NO_PSEUDO_HDR|ACE_RCB_VLAN_FLAG;
1282 for (i = 0; i < RX_MINI_RING_ENTRIES; i++)
1283 ap->rx_mini_ring[i].flags =
1284 BD_FLG_TCP_UDP_SUM | BD_FLG_MINI;
1286 set_aceaddr(&info->rx_mini_ctrl.rngptr, 0);
1287 info->rx_mini_ctrl.flags = RCB_FLG_RNG_DISABLE;
1288 info->rx_mini_ctrl.max_len = 0;
1291 ap->rx_mini_skbprd = 0;
1292 atomic_set(&ap->cur_mini_bufs, 0);
1294 set_aceaddr(&info->rx_return_ctrl.rngptr,
1295 (ap->rx_ring_base_dma +
1296 (sizeof(struct rx_desc) *
1297 (RX_STD_RING_ENTRIES +
1298 RX_JUMBO_RING_ENTRIES +
1299 RX_MINI_RING_ENTRIES))));
1300 info->rx_return_ctrl.flags = 0;
1301 info->rx_return_ctrl.max_len = RX_RETURN_RING_ENTRIES;
1303 memset(ap->rx_return_ring, 0,
1304 RX_RETURN_RING_ENTRIES * sizeof(struct rx_desc));
1306 set_aceaddr(&info->rx_ret_prd_ptr, ap->rx_ret_prd_dma);
1307 *(ap->rx_ret_prd) = 0;
1309 writel(TX_RING_BASE, ®s->WinBase);
1311 if (ACE_IS_TIGON_I(ap)) {
1312 ap->tx_ring = (__force struct tx_desc *) regs->Window;
1313 for (i = 0; i < (TIGON_I_TX_RING_ENTRIES
1314 * sizeof(struct tx_desc)) / sizeof(u32); i++)
1315 writel(0, (__force void __iomem *)ap->tx_ring + i * 4);
1317 set_aceaddr(&info->tx_ctrl.rngptr, TX_RING_BASE);
1319 memset(ap->tx_ring, 0,
1320 MAX_TX_RING_ENTRIES * sizeof(struct tx_desc));
1322 set_aceaddr(&info->tx_ctrl.rngptr, ap->tx_ring_dma);
1325 info->tx_ctrl.max_len = ACE_TX_RING_ENTRIES(ap);
1326 tmp = RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | ACE_RCB_VLAN_FLAG;
1329 * The Tigon I does not like having the TX ring in host memory ;-(
1331 if (!ACE_IS_TIGON_I(ap))
1332 tmp |= RCB_FLG_TX_HOST_RING;
1333 #if TX_COAL_INTS_ONLY
1334 tmp |= RCB_FLG_COAL_INT_ONLY;
1336 info->tx_ctrl.flags = tmp;
1338 set_aceaddr(&info->tx_csm_ptr, ap->tx_csm_dma);
1341 * Potential item for tuning parameter
1344 writel(DMA_THRESH_16W, ®s->DmaReadCfg);
1345 writel(DMA_THRESH_16W, ®s->DmaWriteCfg);
1347 writel(DMA_THRESH_8W, ®s->DmaReadCfg);
1348 writel(DMA_THRESH_8W, ®s->DmaWriteCfg);
1351 writel(0, ®s->MaskInt);
1352 writel(1, ®s->IfIdx);
1355 * McKinley boxes do not like us fiddling with AssistState
1358 writel(1, ®s->AssistState);
1361 writel(DEF_STAT, ®s->TuneStatTicks);
1362 writel(DEF_TRACE, ®s->TuneTrace);
1364 ace_set_rxtx_parms(dev, 0);
1366 if (board_idx == BOARD_IDX_OVERFLOW) {
1367 printk(KERN_WARNING "%s: more than %i NICs detected, "
1368 "ignoring module parameters!\n",
1369 ap->name, ACE_MAX_MOD_PARMS);
1370 } else if (board_idx >= 0) {
1371 if (tx_coal_tick[board_idx])
1372 writel(tx_coal_tick[board_idx],
1373 ®s->TuneTxCoalTicks);
1374 if (max_tx_desc[board_idx])
1375 writel(max_tx_desc[board_idx], ®s->TuneMaxTxDesc);
1377 if (rx_coal_tick[board_idx])
1378 writel(rx_coal_tick[board_idx],
1379 ®s->TuneRxCoalTicks);
1380 if (max_rx_desc[board_idx])
1381 writel(max_rx_desc[board_idx], ®s->TuneMaxRxDesc);
1383 if (trace[board_idx])
1384 writel(trace[board_idx], ®s->TuneTrace);
1386 if ((tx_ratio[board_idx] > 0) && (tx_ratio[board_idx] < 64))
1387 writel(tx_ratio[board_idx], ®s->TxBufRat);
1391 * Default link parameters
1393 tmp = LNK_ENABLE | LNK_FULL_DUPLEX | LNK_1000MB | LNK_100MB |
1394 LNK_10MB | LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL | LNK_NEGOTIATE;
1395 if(ap->version >= 2)
1396 tmp |= LNK_TX_FLOW_CTL_Y;
1399 * Override link default parameters
1401 if ((board_idx >= 0) && link_state[board_idx]) {
1402 int option = link_state[board_idx];
1406 if (option & 0x01) {
1407 printk(KERN_INFO "%s: Setting half duplex link\n",
1409 tmp &= ~LNK_FULL_DUPLEX;
1412 tmp &= ~LNK_NEGOTIATE;
1419 if ((option & 0x70) == 0) {
1420 printk(KERN_WARNING "%s: No media speed specified, "
1421 "forcing auto negotiation\n", ap->name);
1422 tmp |= LNK_NEGOTIATE | LNK_1000MB |
1423 LNK_100MB | LNK_10MB;
1425 if ((option & 0x100) == 0)
1426 tmp |= LNK_NEG_FCTL;
1428 printk(KERN_INFO "%s: Disabling flow control "
1429 "negotiation\n", ap->name);
1431 tmp |= LNK_RX_FLOW_CTL_Y;
1432 if ((option & 0x400) && (ap->version >= 2)) {
1433 printk(KERN_INFO "%s: Enabling TX flow control\n",
1435 tmp |= LNK_TX_FLOW_CTL_Y;
1440 writel(tmp, ®s->TuneLink);
1441 if (ap->version >= 2)
1442 writel(tmp, ®s->TuneFastLink);
1444 if (ACE_IS_TIGON_I(ap))
1445 writel(tigonFwStartAddr, ®s->Pc);
1446 if (ap->version == 2)
1447 writel(tigon2FwStartAddr, ®s->Pc);
1449 writel(0, ®s->Mb0Lo);
1452 * Set tx_csm before we start receiving interrupts, otherwise
1453 * the interrupt handler might think it is supposed to process
1454 * tx ints before we are up and running, which may cause a null
1455 * pointer access in the int handler.
1458 ap->tx_prd = *(ap->tx_csm) = ap->tx_ret_csm = 0;
1461 ace_set_txprd(regs, ap, 0);
1462 writel(0, ®s->RxRetCsm);
1465 * Enable DMA engine now.
1466 * If we do this sooner, Mckinley box pukes.
1467 * I assume it's because Tigon II DMA engine wants to check
1468 * *something* even before the CPU is started.
1470 writel(1, ®s->AssistState); /* enable DMA */
1475 writel(readl(®s->CpuCtrl) & ~(CPU_HALT|CPU_TRACE), ®s->CpuCtrl);
1476 readl(®s->CpuCtrl);
1479 * Wait for the firmware to spin up - max 3 seconds.
1481 myjif = jiffies + 3 * HZ;
1482 while (time_before(jiffies, myjif) && !ap->fw_running)
1485 if (!ap->fw_running) {
1486 printk(KERN_ERR "%s: Firmware NOT running!\n", ap->name);
1489 writel(readl(®s->CpuCtrl) | CPU_HALT, ®s->CpuCtrl);
1490 readl(®s->CpuCtrl);
1492 /* aman@sgi.com - account for badly behaving firmware/NIC:
1493 * - have observed that the NIC may continue to generate
1494 * interrupts for some reason; attempt to stop it - halt
1495 * second CPU for Tigon II cards, and also clear Mb0
1496 * - if we're a module, we'll fail to load if this was
1497 * the only GbE card in the system => if the kernel does
1498 * see an interrupt from the NIC, code to handle it is
1499 * gone and OOps! - so free_irq also
1501 if (ap->version >= 2)
1502 writel(readl(®s->CpuBCtrl) | CPU_HALT,
1504 writel(0, ®s->Mb0Lo);
1505 readl(®s->Mb0Lo);
1512 * We load the ring here as there seem to be no way to tell the
1513 * firmware to wipe the ring without re-initializing it.
1515 if (!test_and_set_bit(0, &ap->std_refill_busy))
1516 ace_load_std_rx_ring(ap, RX_RING_SIZE);
1518 printk(KERN_ERR "%s: Someone is busy refilling the RX ring\n",
1520 if (ap->version >= 2) {
1521 if (!test_and_set_bit(0, &ap->mini_refill_busy))
1522 ace_load_mini_rx_ring(ap, RX_MINI_SIZE);
1524 printk(KERN_ERR "%s: Someone is busy refilling "
1525 "the RX mini ring\n", ap->name);
1530 ace_init_cleanup(dev);
1535 static void ace_set_rxtx_parms(struct net_device *dev, int jumbo)
1537 struct ace_private *ap = netdev_priv(dev);
1538 struct ace_regs __iomem *regs = ap->regs;
1539 int board_idx = ap->board_idx;
1541 if (board_idx >= 0) {
1543 if (!tx_coal_tick[board_idx])
1544 writel(DEF_TX_COAL, ®s->TuneTxCoalTicks);
1545 if (!max_tx_desc[board_idx])
1546 writel(DEF_TX_MAX_DESC, ®s->TuneMaxTxDesc);
1547 if (!rx_coal_tick[board_idx])
1548 writel(DEF_RX_COAL, ®s->TuneRxCoalTicks);
1549 if (!max_rx_desc[board_idx])
1550 writel(DEF_RX_MAX_DESC, ®s->TuneMaxRxDesc);
1551 if (!tx_ratio[board_idx])
1552 writel(DEF_TX_RATIO, ®s->TxBufRat);
1554 if (!tx_coal_tick[board_idx])
1555 writel(DEF_JUMBO_TX_COAL,
1556 ®s->TuneTxCoalTicks);
1557 if (!max_tx_desc[board_idx])
1558 writel(DEF_JUMBO_TX_MAX_DESC,
1559 ®s->TuneMaxTxDesc);
1560 if (!rx_coal_tick[board_idx])
1561 writel(DEF_JUMBO_RX_COAL,
1562 ®s->TuneRxCoalTicks);
1563 if (!max_rx_desc[board_idx])
1564 writel(DEF_JUMBO_RX_MAX_DESC,
1565 ®s->TuneMaxRxDesc);
1566 if (!tx_ratio[board_idx])
1567 writel(DEF_JUMBO_TX_RATIO, ®s->TxBufRat);
1573 static void ace_watchdog(struct net_device *data)
1575 struct net_device *dev = data;
1576 struct ace_private *ap = netdev_priv(dev);
1577 struct ace_regs __iomem *regs = ap->regs;
1580 * We haven't received a stats update event for more than 2.5
1581 * seconds and there is data in the transmit queue, thus we
1582 * asume the card is stuck.
1584 if (*ap->tx_csm != ap->tx_ret_csm) {
1585 printk(KERN_WARNING "%s: Transmitter is stuck, %08x\n",
1586 dev->name, (unsigned int)readl(®s->HostCtrl));
1587 /* This can happen due to ieee flow control. */
1589 printk(KERN_DEBUG "%s: BUG... transmitter died. Kicking it.\n",
1592 netif_wake_queue(dev);
1598 static void ace_tasklet(unsigned long dev)
1600 struct ace_private *ap = netdev_priv((struct net_device *)dev);
1603 cur_size = atomic_read(&ap->cur_rx_bufs);
1604 if ((cur_size < RX_LOW_STD_THRES) &&
1605 !test_and_set_bit(0, &ap->std_refill_busy)) {
1607 printk("refilling buffers (current %i)\n", cur_size);
1609 ace_load_std_rx_ring(ap, RX_RING_SIZE - cur_size);
1612 if (ap->version >= 2) {
1613 cur_size = atomic_read(&ap->cur_mini_bufs);
1614 if ((cur_size < RX_LOW_MINI_THRES) &&
1615 !test_and_set_bit(0, &ap->mini_refill_busy)) {
1617 printk("refilling mini buffers (current %i)\n",
1620 ace_load_mini_rx_ring(ap, RX_MINI_SIZE - cur_size);
1624 cur_size = atomic_read(&ap->cur_jumbo_bufs);
1625 if (ap->jumbo && (cur_size < RX_LOW_JUMBO_THRES) &&
1626 !test_and_set_bit(0, &ap->jumbo_refill_busy)) {
1628 printk("refilling jumbo buffers (current %i)\n", cur_size);
1630 ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE - cur_size);
1632 ap->tasklet_pending = 0;
1637 * Copy the contents of the NIC's trace buffer to kernel memory.
1639 static void ace_dump_trace(struct ace_private *ap)
1643 if (!(ap->trace_buf = kmalloc(ACE_TRACE_SIZE, GFP_KERNEL)))
1650 * Load the standard rx ring.
1652 * Loading rings is safe without holding the spin lock since this is
1653 * done only before the device is enabled, thus no interrupts are
1654 * generated and by the interrupt handler/tasklet handler.
1656 static void ace_load_std_rx_ring(struct ace_private *ap, int nr_bufs)
1658 struct ace_regs __iomem *regs = ap->regs;
1662 prefetchw(&ap->cur_rx_bufs);
1664 idx = ap->rx_std_skbprd;
1666 for (i = 0; i < nr_bufs; i++) {
1667 struct sk_buff *skb;
1671 skb = alloc_skb(ACE_STD_BUFSIZE + NET_IP_ALIGN, GFP_ATOMIC);
1675 skb_reserve(skb, NET_IP_ALIGN);
1676 mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1677 offset_in_page(skb->data),
1679 PCI_DMA_FROMDEVICE);
1680 ap->skb->rx_std_skbuff[idx].skb = skb;
1681 pci_unmap_addr_set(&ap->skb->rx_std_skbuff[idx],
1684 rd = &ap->rx_std_ring[idx];
1685 set_aceaddr(&rd->addr, mapping);
1686 rd->size = ACE_STD_BUFSIZE;
1688 idx = (idx + 1) % RX_STD_RING_ENTRIES;
1694 atomic_add(i, &ap->cur_rx_bufs);
1695 ap->rx_std_skbprd = idx;
1697 if (ACE_IS_TIGON_I(ap)) {
1699 cmd.evt = C_SET_RX_PRD_IDX;
1701 cmd.idx = ap->rx_std_skbprd;
1702 ace_issue_cmd(regs, &cmd);
1704 writel(idx, ®s->RxStdPrd);
1709 clear_bit(0, &ap->std_refill_busy);
1713 printk(KERN_INFO "Out of memory when allocating "
1714 "standard receive buffers\n");
1719 static void ace_load_mini_rx_ring(struct ace_private *ap, int nr_bufs)
1721 struct ace_regs __iomem *regs = ap->regs;
1724 prefetchw(&ap->cur_mini_bufs);
1726 idx = ap->rx_mini_skbprd;
1727 for (i = 0; i < nr_bufs; i++) {
1728 struct sk_buff *skb;
1732 skb = alloc_skb(ACE_MINI_BUFSIZE + NET_IP_ALIGN, GFP_ATOMIC);
1736 skb_reserve(skb, NET_IP_ALIGN);
1737 mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1738 offset_in_page(skb->data),
1740 PCI_DMA_FROMDEVICE);
1741 ap->skb->rx_mini_skbuff[idx].skb = skb;
1742 pci_unmap_addr_set(&ap->skb->rx_mini_skbuff[idx],
1745 rd = &ap->rx_mini_ring[idx];
1746 set_aceaddr(&rd->addr, mapping);
1747 rd->size = ACE_MINI_BUFSIZE;
1749 idx = (idx + 1) % RX_MINI_RING_ENTRIES;
1755 atomic_add(i, &ap->cur_mini_bufs);
1757 ap->rx_mini_skbprd = idx;
1759 writel(idx, ®s->RxMiniPrd);
1763 clear_bit(0, &ap->mini_refill_busy);
1766 printk(KERN_INFO "Out of memory when allocating "
1767 "mini receive buffers\n");
1773 * Load the jumbo rx ring, this may happen at any time if the MTU
1774 * is changed to a value > 1500.
1776 static void ace_load_jumbo_rx_ring(struct ace_private *ap, int nr_bufs)
1778 struct ace_regs __iomem *regs = ap->regs;
1781 idx = ap->rx_jumbo_skbprd;
1783 for (i = 0; i < nr_bufs; i++) {
1784 struct sk_buff *skb;
1788 skb = alloc_skb(ACE_JUMBO_BUFSIZE + NET_IP_ALIGN, GFP_ATOMIC);
1792 skb_reserve(skb, NET_IP_ALIGN);
1793 mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1794 offset_in_page(skb->data),
1796 PCI_DMA_FROMDEVICE);
1797 ap->skb->rx_jumbo_skbuff[idx].skb = skb;
1798 pci_unmap_addr_set(&ap->skb->rx_jumbo_skbuff[idx],
1801 rd = &ap->rx_jumbo_ring[idx];
1802 set_aceaddr(&rd->addr, mapping);
1803 rd->size = ACE_JUMBO_BUFSIZE;
1805 idx = (idx + 1) % RX_JUMBO_RING_ENTRIES;
1811 atomic_add(i, &ap->cur_jumbo_bufs);
1812 ap->rx_jumbo_skbprd = idx;
1814 if (ACE_IS_TIGON_I(ap)) {
1816 cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1818 cmd.idx = ap->rx_jumbo_skbprd;
1819 ace_issue_cmd(regs, &cmd);
1821 writel(idx, ®s->RxJumboPrd);
1826 clear_bit(0, &ap->jumbo_refill_busy);
1829 if (net_ratelimit())
1830 printk(KERN_INFO "Out of memory when allocating "
1831 "jumbo receive buffers\n");
1837 * All events are considered to be slow (RX/TX ints do not generate
1838 * events) and are handled here, outside the main interrupt handler,
1839 * to reduce the size of the handler.
1841 static u32 ace_handle_event(struct net_device *dev, u32 evtcsm, u32 evtprd)
1843 struct ace_private *ap;
1845 ap = netdev_priv(dev);
1847 while (evtcsm != evtprd) {
1848 switch (ap->evt_ring[evtcsm].evt) {
1850 printk(KERN_INFO "%s: Firmware up and running\n",
1855 case E_STATS_UPDATED:
1859 u16 code = ap->evt_ring[evtcsm].code;
1863 u32 state = readl(&ap->regs->GigLnkState);
1864 printk(KERN_WARNING "%s: Optical link UP "
1865 "(%s Duplex, Flow Control: %s%s)\n",
1867 state & LNK_FULL_DUPLEX ? "Full":"Half",
1868 state & LNK_TX_FLOW_CTL_Y ? "TX " : "",
1869 state & LNK_RX_FLOW_CTL_Y ? "RX" : "");
1873 printk(KERN_WARNING "%s: Optical link DOWN\n",
1876 case E_C_LINK_10_100:
1877 printk(KERN_WARNING "%s: 10/100BaseT link "
1881 printk(KERN_ERR "%s: Unknown optical link "
1882 "state %02x\n", ap->name, code);
1887 switch(ap->evt_ring[evtcsm].code) {
1888 case E_C_ERR_INVAL_CMD:
1889 printk(KERN_ERR "%s: invalid command error\n",
1892 case E_C_ERR_UNIMP_CMD:
1893 printk(KERN_ERR "%s: unimplemented command "
1894 "error\n", ap->name);
1896 case E_C_ERR_BAD_CFG:
1897 printk(KERN_ERR "%s: bad config error\n",
1901 printk(KERN_ERR "%s: unknown error %02x\n",
1902 ap->name, ap->evt_ring[evtcsm].code);
1905 case E_RESET_JUMBO_RNG:
1908 for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
1909 if (ap->skb->rx_jumbo_skbuff[i].skb) {
1910 ap->rx_jumbo_ring[i].size = 0;
1911 set_aceaddr(&ap->rx_jumbo_ring[i].addr, 0);
1912 dev_kfree_skb(ap->skb->rx_jumbo_skbuff[i].skb);
1913 ap->skb->rx_jumbo_skbuff[i].skb = NULL;
1917 if (ACE_IS_TIGON_I(ap)) {
1919 cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1922 ace_issue_cmd(ap->regs, &cmd);
1924 writel(0, &((ap->regs)->RxJumboPrd));
1929 ap->rx_jumbo_skbprd = 0;
1930 printk(KERN_INFO "%s: Jumbo ring flushed\n",
1932 clear_bit(0, &ap->jumbo_refill_busy);
1936 printk(KERN_ERR "%s: Unhandled event 0x%02x\n",
1937 ap->name, ap->evt_ring[evtcsm].evt);
1939 evtcsm = (evtcsm + 1) % EVT_RING_ENTRIES;
1946 static void ace_rx_int(struct net_device *dev, u32 rxretprd, u32 rxretcsm)
1948 struct ace_private *ap = netdev_priv(dev);
1950 int mini_count = 0, std_count = 0;
1954 prefetchw(&ap->cur_rx_bufs);
1955 prefetchw(&ap->cur_mini_bufs);
1957 while (idx != rxretprd) {
1958 struct ring_info *rip;
1959 struct sk_buff *skb;
1960 struct rx_desc *rxdesc, *retdesc;
1962 int bd_flags, desc_type, mapsize;
1966 /* make sure the rx descriptor isn't read before rxretprd */
1967 if (idx == rxretcsm)
1970 retdesc = &ap->rx_return_ring[idx];
1971 skbidx = retdesc->idx;
1972 bd_flags = retdesc->flags;
1973 desc_type = bd_flags & (BD_FLG_JUMBO | BD_FLG_MINI);
1977 * Normal frames do not have any flags set
1979 * Mini and normal frames arrive frequently,
1980 * so use a local counter to avoid doing
1981 * atomic operations for each packet arriving.
1984 rip = &ap->skb->rx_std_skbuff[skbidx];
1985 mapsize = ACE_STD_BUFSIZE;
1986 rxdesc = &ap->rx_std_ring[skbidx];
1990 rip = &ap->skb->rx_jumbo_skbuff[skbidx];
1991 mapsize = ACE_JUMBO_BUFSIZE;
1992 rxdesc = &ap->rx_jumbo_ring[skbidx];
1993 atomic_dec(&ap->cur_jumbo_bufs);
1996 rip = &ap->skb->rx_mini_skbuff[skbidx];
1997 mapsize = ACE_MINI_BUFSIZE;
1998 rxdesc = &ap->rx_mini_ring[skbidx];
2002 printk(KERN_INFO "%s: unknown frame type (0x%02x) "
2003 "returned by NIC\n", dev->name,
2010 pci_unmap_page(ap->pdev,
2011 pci_unmap_addr(rip, mapping),
2013 PCI_DMA_FROMDEVICE);
2014 skb_put(skb, retdesc->size);
2019 csum = retdesc->tcp_udp_csum;
2021 skb->protocol = eth_type_trans(skb, dev);
2024 * Instead of forcing the poor tigon mips cpu to calculate
2025 * pseudo hdr checksum, we do this ourselves.
2027 if (bd_flags & BD_FLG_TCP_UDP_SUM) {
2028 skb->csum = htons(csum);
2029 skb->ip_summed = CHECKSUM_COMPLETE;
2031 skb->ip_summed = CHECKSUM_NONE;
2036 if (ap->vlgrp && (bd_flags & BD_FLG_VLAN_TAG)) {
2037 vlan_hwaccel_rx(skb, ap->vlgrp, retdesc->vlan);
2042 dev->stats.rx_packets++;
2043 dev->stats.rx_bytes += retdesc->size;
2045 idx = (idx + 1) % RX_RETURN_RING_ENTRIES;
2048 atomic_sub(std_count, &ap->cur_rx_bufs);
2049 if (!ACE_IS_TIGON_I(ap))
2050 atomic_sub(mini_count, &ap->cur_mini_bufs);
2054 * According to the documentation RxRetCsm is obsolete with
2055 * the 12.3.x Firmware - my Tigon I NICs seem to disagree!
2057 if (ACE_IS_TIGON_I(ap)) {
2058 writel(idx, &ap->regs->RxRetCsm);
2069 static inline void ace_tx_int(struct net_device *dev,
2072 struct ace_private *ap = netdev_priv(dev);
2075 struct sk_buff *skb;
2077 struct tx_ring_info *info;
2079 info = ap->skb->tx_skbuff + idx;
2081 mapping = pci_unmap_addr(info, mapping);
2084 pci_unmap_page(ap->pdev, mapping,
2085 pci_unmap_len(info, maplen),
2087 pci_unmap_addr_set(info, mapping, 0);
2091 dev->stats.tx_packets++;
2092 dev->stats.tx_bytes += skb->len;
2093 dev_kfree_skb_irq(skb);
2097 idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2098 } while (idx != txcsm);
2100 if (netif_queue_stopped(dev))
2101 netif_wake_queue(dev);
2104 ap->tx_ret_csm = txcsm;
2106 /* So... tx_ret_csm is advanced _after_ check for device wakeup.
2108 * We could try to make it before. In this case we would get
2109 * the following race condition: hard_start_xmit on other cpu
2110 * enters after we advanced tx_ret_csm and fills space,
2111 * which we have just freed, so that we make illegal device wakeup.
2112 * There is no good way to workaround this (at entry
2113 * to ace_start_xmit detects this condition and prevents
2114 * ring corruption, but it is not a good workaround.)
2116 * When tx_ret_csm is advanced after, we wake up device _only_
2117 * if we really have some space in ring (though the core doing
2118 * hard_start_xmit can see full ring for some period and has to
2119 * synchronize.) Superb.
2120 * BUT! We get another subtle race condition. hard_start_xmit
2121 * may think that ring is full between wakeup and advancing
2122 * tx_ret_csm and will stop device instantly! It is not so bad.
2123 * We are guaranteed that there is something in ring, so that
2124 * the next irq will resume transmission. To speedup this we could
2125 * mark descriptor, which closes ring with BD_FLG_COAL_NOW
2126 * (see ace_start_xmit).
2128 * Well, this dilemma exists in all lock-free devices.
2129 * We, following scheme used in drivers by Donald Becker,
2130 * select the least dangerous.
2136 static irqreturn_t ace_interrupt(int irq, void *dev_id)
2138 struct net_device *dev = (struct net_device *)dev_id;
2139 struct ace_private *ap = netdev_priv(dev);
2140 struct ace_regs __iomem *regs = ap->regs;
2142 u32 txcsm, rxretcsm, rxretprd;
2146 * In case of PCI shared interrupts or spurious interrupts,
2147 * we want to make sure it is actually our interrupt before
2148 * spending any time in here.
2150 if (!(readl(®s->HostCtrl) & IN_INT))
2154 * ACK intr now. Otherwise we will lose updates to rx_ret_prd,
2155 * which happened _after_ rxretprd = *ap->rx_ret_prd; but before
2156 * writel(0, ®s->Mb0Lo).
2158 * "IRQ avoidance" recommended in docs applies to IRQs served
2159 * threads and it is wrong even for that case.
2161 writel(0, ®s->Mb0Lo);
2162 readl(®s->Mb0Lo);
2165 * There is no conflict between transmit handling in
2166 * start_xmit and receive processing, thus there is no reason
2167 * to take a spin lock for RX handling. Wait until we start
2168 * working on the other stuff - hey we don't need a spin lock
2171 rxretprd = *ap->rx_ret_prd;
2172 rxretcsm = ap->cur_rx;
2174 if (rxretprd != rxretcsm)
2175 ace_rx_int(dev, rxretprd, rxretcsm);
2177 txcsm = *ap->tx_csm;
2178 idx = ap->tx_ret_csm;
2182 * If each skb takes only one descriptor this check degenerates
2183 * to identity, because new space has just been opened.
2184 * But if skbs are fragmented we must check that this index
2185 * update releases enough of space, otherwise we just
2186 * wait for device to make more work.
2188 if (!tx_ring_full(ap, txcsm, ap->tx_prd))
2189 ace_tx_int(dev, txcsm, idx);
2192 evtcsm = readl(®s->EvtCsm);
2193 evtprd = *ap->evt_prd;
2195 if (evtcsm != evtprd) {
2196 evtcsm = ace_handle_event(dev, evtcsm, evtprd);
2197 writel(evtcsm, ®s->EvtCsm);
2201 * This has to go last in the interrupt handler and run with
2202 * the spin lock released ... what lock?
2204 if (netif_running(dev)) {
2206 int run_tasklet = 0;
2208 cur_size = atomic_read(&ap->cur_rx_bufs);
2209 if (cur_size < RX_LOW_STD_THRES) {
2210 if ((cur_size < RX_PANIC_STD_THRES) &&
2211 !test_and_set_bit(0, &ap->std_refill_busy)) {
2213 printk("low on std buffers %i\n", cur_size);
2215 ace_load_std_rx_ring(ap,
2216 RX_RING_SIZE - cur_size);
2221 if (!ACE_IS_TIGON_I(ap)) {
2222 cur_size = atomic_read(&ap->cur_mini_bufs);
2223 if (cur_size < RX_LOW_MINI_THRES) {
2224 if ((cur_size < RX_PANIC_MINI_THRES) &&
2225 !test_and_set_bit(0,
2226 &ap->mini_refill_busy)) {
2228 printk("low on mini buffers %i\n",
2231 ace_load_mini_rx_ring(ap, RX_MINI_SIZE - cur_size);
2238 cur_size = atomic_read(&ap->cur_jumbo_bufs);
2239 if (cur_size < RX_LOW_JUMBO_THRES) {
2240 if ((cur_size < RX_PANIC_JUMBO_THRES) &&
2241 !test_and_set_bit(0,
2242 &ap->jumbo_refill_busy)){
2244 printk("low on jumbo buffers %i\n",
2247 ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE - cur_size);
2252 if (run_tasklet && !ap->tasklet_pending) {
2253 ap->tasklet_pending = 1;
2254 tasklet_schedule(&ap->ace_tasklet);
2263 static void ace_vlan_rx_register(struct net_device *dev, struct vlan_group *grp)
2265 struct ace_private *ap = netdev_priv(dev);
2266 unsigned long flags;
2268 local_irq_save(flags);
2273 ace_unmask_irq(dev);
2274 local_irq_restore(flags);
2276 #endif /* ACENIC_DO_VLAN */
2279 static int ace_open(struct net_device *dev)
2281 struct ace_private *ap = netdev_priv(dev);
2282 struct ace_regs __iomem *regs = ap->regs;
2285 if (!(ap->fw_running)) {
2286 printk(KERN_WARNING "%s: Firmware not running!\n", dev->name);
2290 writel(dev->mtu + ETH_HLEN + 4, ®s->IfMtu);
2292 cmd.evt = C_CLEAR_STATS;
2295 ace_issue_cmd(regs, &cmd);
2297 cmd.evt = C_HOST_STATE;
2298 cmd.code = C_C_STACK_UP;
2300 ace_issue_cmd(regs, &cmd);
2303 !test_and_set_bit(0, &ap->jumbo_refill_busy))
2304 ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE);
2306 if (dev->flags & IFF_PROMISC) {
2307 cmd.evt = C_SET_PROMISC_MODE;
2308 cmd.code = C_C_PROMISC_ENABLE;
2310 ace_issue_cmd(regs, &cmd);
2318 cmd.evt = C_LNK_NEGOTIATION;
2321 ace_issue_cmd(regs, &cmd);
2324 netif_start_queue(dev);
2327 * Setup the bottom half rx ring refill handler
2329 tasklet_init(&ap->ace_tasklet, ace_tasklet, (unsigned long)dev);
2334 static int ace_close(struct net_device *dev)
2336 struct ace_private *ap = netdev_priv(dev);
2337 struct ace_regs __iomem *regs = ap->regs;
2339 unsigned long flags;
2343 * Without (or before) releasing irq and stopping hardware, this
2344 * is an absolute non-sense, by the way. It will be reset instantly
2347 netif_stop_queue(dev);
2351 cmd.evt = C_SET_PROMISC_MODE;
2352 cmd.code = C_C_PROMISC_DISABLE;
2354 ace_issue_cmd(regs, &cmd);
2358 cmd.evt = C_HOST_STATE;
2359 cmd.code = C_C_STACK_DOWN;
2361 ace_issue_cmd(regs, &cmd);
2363 tasklet_kill(&ap->ace_tasklet);
2366 * Make sure one CPU is not processing packets while
2367 * buffers are being released by another.
2370 local_irq_save(flags);
2373 for (i = 0; i < ACE_TX_RING_ENTRIES(ap); i++) {
2374 struct sk_buff *skb;
2376 struct tx_ring_info *info;
2378 info = ap->skb->tx_skbuff + i;
2380 mapping = pci_unmap_addr(info, mapping);
2383 if (ACE_IS_TIGON_I(ap)) {
2384 /* NB: TIGON_1 is special, tx_ring is in io space */
2385 struct tx_desc __iomem *tx;
2386 tx = (__force struct tx_desc __iomem *) &ap->tx_ring[i];
2387 writel(0, &tx->addr.addrhi);
2388 writel(0, &tx->addr.addrlo);
2389 writel(0, &tx->flagsize);
2391 memset(ap->tx_ring + i, 0,
2392 sizeof(struct tx_desc));
2393 pci_unmap_page(ap->pdev, mapping,
2394 pci_unmap_len(info, maplen),
2396 pci_unmap_addr_set(info, mapping, 0);
2405 cmd.evt = C_RESET_JUMBO_RNG;
2408 ace_issue_cmd(regs, &cmd);
2411 ace_unmask_irq(dev);
2412 local_irq_restore(flags);
2418 static inline dma_addr_t
2419 ace_map_tx_skb(struct ace_private *ap, struct sk_buff *skb,
2420 struct sk_buff *tail, u32 idx)
2423 struct tx_ring_info *info;
2425 mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
2426 offset_in_page(skb->data),
2427 skb->len, PCI_DMA_TODEVICE);
2429 info = ap->skb->tx_skbuff + idx;
2431 pci_unmap_addr_set(info, mapping, mapping);
2432 pci_unmap_len_set(info, maplen, skb->len);
2438 ace_load_tx_bd(struct ace_private *ap, struct tx_desc *desc, u64 addr,
2439 u32 flagsize, u32 vlan_tag)
2441 #if !USE_TX_COAL_NOW
2442 flagsize &= ~BD_FLG_COAL_NOW;
2445 if (ACE_IS_TIGON_I(ap)) {
2446 struct tx_desc __iomem *io = (__force struct tx_desc __iomem *) desc;
2447 writel(addr >> 32, &io->addr.addrhi);
2448 writel(addr & 0xffffffff, &io->addr.addrlo);
2449 writel(flagsize, &io->flagsize);
2451 writel(vlan_tag, &io->vlanres);
2454 desc->addr.addrhi = addr >> 32;
2455 desc->addr.addrlo = addr;
2456 desc->flagsize = flagsize;
2458 desc->vlanres = vlan_tag;
2464 static int ace_start_xmit(struct sk_buff *skb, struct net_device *dev)
2466 struct ace_private *ap = netdev_priv(dev);
2467 struct ace_regs __iomem *regs = ap->regs;
2468 struct tx_desc *desc;
2470 unsigned long maxjiff = jiffies + 3*HZ;
2475 if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2478 if (!skb_shinfo(skb)->nr_frags) {
2482 mapping = ace_map_tx_skb(ap, skb, skb, idx);
2483 flagsize = (skb->len << 16) | (BD_FLG_END);
2484 if (skb->ip_summed == CHECKSUM_PARTIAL)
2485 flagsize |= BD_FLG_TCP_UDP_SUM;
2487 if (vlan_tx_tag_present(skb)) {
2488 flagsize |= BD_FLG_VLAN_TAG;
2489 vlan_tag = vlan_tx_tag_get(skb);
2492 desc = ap->tx_ring + idx;
2493 idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2495 /* Look at ace_tx_int for explanations. */
2496 if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2497 flagsize |= BD_FLG_COAL_NOW;
2499 ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2505 mapping = ace_map_tx_skb(ap, skb, NULL, idx);
2506 flagsize = (skb_headlen(skb) << 16);
2507 if (skb->ip_summed == CHECKSUM_PARTIAL)
2508 flagsize |= BD_FLG_TCP_UDP_SUM;
2510 if (vlan_tx_tag_present(skb)) {
2511 flagsize |= BD_FLG_VLAN_TAG;
2512 vlan_tag = vlan_tx_tag_get(skb);
2516 ace_load_tx_bd(ap, ap->tx_ring + idx, mapping, flagsize, vlan_tag);
2518 idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2520 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2521 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2522 struct tx_ring_info *info;
2525 info = ap->skb->tx_skbuff + idx;
2526 desc = ap->tx_ring + idx;
2528 mapping = pci_map_page(ap->pdev, frag->page,
2529 frag->page_offset, frag->size,
2532 flagsize = (frag->size << 16);
2533 if (skb->ip_summed == CHECKSUM_PARTIAL)
2534 flagsize |= BD_FLG_TCP_UDP_SUM;
2535 idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2537 if (i == skb_shinfo(skb)->nr_frags - 1) {
2538 flagsize |= BD_FLG_END;
2539 if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2540 flagsize |= BD_FLG_COAL_NOW;
2543 * Only the last fragment frees
2550 pci_unmap_addr_set(info, mapping, mapping);
2551 pci_unmap_len_set(info, maplen, frag->size);
2552 ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2558 ace_set_txprd(regs, ap, idx);
2560 if (flagsize & BD_FLG_COAL_NOW) {
2561 netif_stop_queue(dev);
2564 * A TX-descriptor producer (an IRQ) might have gotten
2565 * inbetween, making the ring free again. Since xmit is
2566 * serialized, this is the only situation we have to
2569 if (!tx_ring_full(ap, ap->tx_ret_csm, idx))
2570 netif_wake_queue(dev);
2573 dev->trans_start = jiffies;
2574 return NETDEV_TX_OK;
2578 * This race condition is unavoidable with lock-free drivers.
2579 * We wake up the queue _before_ tx_prd is advanced, so that we can
2580 * enter hard_start_xmit too early, while tx ring still looks closed.
2581 * This happens ~1-4 times per 100000 packets, so that we can allow
2582 * to loop syncing to other CPU. Probably, we need an additional
2583 * wmb() in ace_tx_intr as well.
2585 * Note that this race is relieved by reserving one more entry
2586 * in tx ring than it is necessary (see original non-SG driver).
2587 * However, with SG we need to reserve 2*MAX_SKB_FRAGS+1, which
2588 * is already overkill.
2590 * Alternative is to return with 1 not throttling queue. In this
2591 * case loop becomes longer, no more useful effects.
2593 if (time_before(jiffies, maxjiff)) {
2599 /* The ring is stuck full. */
2600 printk(KERN_WARNING "%s: Transmit ring stuck full\n", dev->name);
2601 return NETDEV_TX_BUSY;
2605 static int ace_change_mtu(struct net_device *dev, int new_mtu)
2607 struct ace_private *ap = netdev_priv(dev);
2608 struct ace_regs __iomem *regs = ap->regs;
2610 if (new_mtu > ACE_JUMBO_MTU)
2613 writel(new_mtu + ETH_HLEN + 4, ®s->IfMtu);
2616 if (new_mtu > ACE_STD_MTU) {
2618 printk(KERN_INFO "%s: Enabling Jumbo frame "
2619 "support\n", dev->name);
2621 if (!test_and_set_bit(0, &ap->jumbo_refill_busy))
2622 ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE);
2623 ace_set_rxtx_parms(dev, 1);
2626 while (test_and_set_bit(0, &ap->jumbo_refill_busy));
2627 ace_sync_irq(dev->irq);
2628 ace_set_rxtx_parms(dev, 0);
2632 cmd.evt = C_RESET_JUMBO_RNG;
2635 ace_issue_cmd(regs, &cmd);
2642 static int ace_get_settings(struct net_device *dev, struct ethtool_cmd *ecmd)
2644 struct ace_private *ap = netdev_priv(dev);
2645 struct ace_regs __iomem *regs = ap->regs;
2648 memset(ecmd, 0, sizeof(struct ethtool_cmd));
2650 (SUPPORTED_10baseT_Half | SUPPORTED_10baseT_Full |
2651 SUPPORTED_100baseT_Half | SUPPORTED_100baseT_Full |
2652 SUPPORTED_1000baseT_Half | SUPPORTED_1000baseT_Full |
2653 SUPPORTED_Autoneg | SUPPORTED_FIBRE);
2655 ecmd->port = PORT_FIBRE;
2656 ecmd->transceiver = XCVR_INTERNAL;
2658 link = readl(®s->GigLnkState);
2659 if (link & LNK_1000MB)
2660 ecmd->speed = SPEED_1000;
2662 link = readl(®s->FastLnkState);
2663 if (link & LNK_100MB)
2664 ecmd->speed = SPEED_100;
2665 else if (link & LNK_10MB)
2666 ecmd->speed = SPEED_10;
2670 if (link & LNK_FULL_DUPLEX)
2671 ecmd->duplex = DUPLEX_FULL;
2673 ecmd->duplex = DUPLEX_HALF;
2675 if (link & LNK_NEGOTIATE)
2676 ecmd->autoneg = AUTONEG_ENABLE;
2678 ecmd->autoneg = AUTONEG_DISABLE;
2682 * Current struct ethtool_cmd is insufficient
2684 ecmd->trace = readl(®s->TuneTrace);
2686 ecmd->txcoal = readl(®s->TuneTxCoalTicks);
2687 ecmd->rxcoal = readl(®s->TuneRxCoalTicks);
2689 ecmd->maxtxpkt = readl(®s->TuneMaxTxDesc);
2690 ecmd->maxrxpkt = readl(®s->TuneMaxRxDesc);
2695 static int ace_set_settings(struct net_device *dev, struct ethtool_cmd *ecmd)
2697 struct ace_private *ap = netdev_priv(dev);
2698 struct ace_regs __iomem *regs = ap->regs;
2701 link = readl(®s->GigLnkState);
2702 if (link & LNK_1000MB)
2705 link = readl(®s->FastLnkState);
2706 if (link & LNK_100MB)
2708 else if (link & LNK_10MB)
2714 link = LNK_ENABLE | LNK_1000MB | LNK_100MB | LNK_10MB |
2715 LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL;
2716 if (!ACE_IS_TIGON_I(ap))
2717 link |= LNK_TX_FLOW_CTL_Y;
2718 if (ecmd->autoneg == AUTONEG_ENABLE)
2719 link |= LNK_NEGOTIATE;
2720 if (ecmd->speed != speed) {
2721 link &= ~(LNK_1000MB | LNK_100MB | LNK_10MB);
2735 if (ecmd->duplex == DUPLEX_FULL)
2736 link |= LNK_FULL_DUPLEX;
2738 if (link != ap->link) {
2740 printk(KERN_INFO "%s: Renegotiating link state\n",
2744 writel(link, ®s->TuneLink);
2745 if (!ACE_IS_TIGON_I(ap))
2746 writel(link, ®s->TuneFastLink);
2749 cmd.evt = C_LNK_NEGOTIATION;
2752 ace_issue_cmd(regs, &cmd);
2757 static void ace_get_drvinfo(struct net_device *dev,
2758 struct ethtool_drvinfo *info)
2760 struct ace_private *ap = netdev_priv(dev);
2762 strlcpy(info->driver, "acenic", sizeof(info->driver));
2763 snprintf(info->version, sizeof(info->version), "%i.%i.%i",
2764 tigonFwReleaseMajor, tigonFwReleaseMinor,
2768 strlcpy(info->bus_info, pci_name(ap->pdev),
2769 sizeof(info->bus_info));
2774 * Set the hardware MAC address.
2776 static int ace_set_mac_addr(struct net_device *dev, void *p)
2778 struct ace_private *ap = netdev_priv(dev);
2779 struct ace_regs __iomem *regs = ap->regs;
2780 struct sockaddr *addr=p;
2784 if(netif_running(dev))
2787 memcpy(dev->dev_addr, addr->sa_data,dev->addr_len);
2789 da = (u8 *)dev->dev_addr;
2791 writel(da[0] << 8 | da[1], ®s->MacAddrHi);
2792 writel((da[2] << 24) | (da[3] << 16) | (da[4] << 8) | da[5],
2795 cmd.evt = C_SET_MAC_ADDR;
2798 ace_issue_cmd(regs, &cmd);
2804 static void ace_set_multicast_list(struct net_device *dev)
2806 struct ace_private *ap = netdev_priv(dev);
2807 struct ace_regs __iomem *regs = ap->regs;
2810 if ((dev->flags & IFF_ALLMULTI) && !(ap->mcast_all)) {
2811 cmd.evt = C_SET_MULTICAST_MODE;
2812 cmd.code = C_C_MCAST_ENABLE;
2814 ace_issue_cmd(regs, &cmd);
2816 } else if (ap->mcast_all) {
2817 cmd.evt = C_SET_MULTICAST_MODE;
2818 cmd.code = C_C_MCAST_DISABLE;
2820 ace_issue_cmd(regs, &cmd);
2824 if ((dev->flags & IFF_PROMISC) && !(ap->promisc)) {
2825 cmd.evt = C_SET_PROMISC_MODE;
2826 cmd.code = C_C_PROMISC_ENABLE;
2828 ace_issue_cmd(regs, &cmd);
2830 }else if (!(dev->flags & IFF_PROMISC) && (ap->promisc)) {
2831 cmd.evt = C_SET_PROMISC_MODE;
2832 cmd.code = C_C_PROMISC_DISABLE;
2834 ace_issue_cmd(regs, &cmd);
2839 * For the time being multicast relies on the upper layers
2840 * filtering it properly. The Firmware does not allow one to
2841 * set the entire multicast list at a time and keeping track of
2842 * it here is going to be messy.
2844 if ((dev->mc_count) && !(ap->mcast_all)) {
2845 cmd.evt = C_SET_MULTICAST_MODE;
2846 cmd.code = C_C_MCAST_ENABLE;
2848 ace_issue_cmd(regs, &cmd);
2849 }else if (!ap->mcast_all) {
2850 cmd.evt = C_SET_MULTICAST_MODE;
2851 cmd.code = C_C_MCAST_DISABLE;
2853 ace_issue_cmd(regs, &cmd);
2858 static struct net_device_stats *ace_get_stats(struct net_device *dev)
2860 struct ace_private *ap = netdev_priv(dev);
2861 struct ace_mac_stats __iomem *mac_stats =
2862 (struct ace_mac_stats __iomem *)ap->regs->Stats;
2864 dev->stats.rx_missed_errors = readl(&mac_stats->drop_space);
2865 dev->stats.multicast = readl(&mac_stats->kept_mc);
2866 dev->stats.collisions = readl(&mac_stats->coll);
2872 static void __devinit ace_copy(struct ace_regs __iomem *regs, void *src,
2875 void __iomem *tdest;
2883 tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2884 min_t(u32, size, ACE_WINDOW_SIZE));
2885 tdest = (void __iomem *) ®s->Window +
2886 (dest & (ACE_WINDOW_SIZE - 1));
2887 writel(dest & ~(ACE_WINDOW_SIZE - 1), ®s->WinBase);
2889 * This requires byte swapping on big endian, however
2890 * writel does that for us
2893 for (i = 0; i < (tsize / 4); i++) {
2894 writel(wsrc[i], tdest + i*4);
2905 static void __devinit ace_clear(struct ace_regs __iomem *regs, u32 dest, int size)
2907 void __iomem *tdest;
2914 tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2915 min_t(u32, size, ACE_WINDOW_SIZE));
2916 tdest = (void __iomem *) ®s->Window +
2917 (dest & (ACE_WINDOW_SIZE - 1));
2918 writel(dest & ~(ACE_WINDOW_SIZE - 1), ®s->WinBase);
2920 for (i = 0; i < (tsize / 4); i++) {
2921 writel(0, tdest + i*4);
2933 * Download the firmware into the SRAM on the NIC
2935 * This operation requires the NIC to be halted and is performed with
2936 * interrupts disabled and with the spinlock hold.
2938 static int __devinit ace_load_firmware(struct net_device *dev)
2940 struct ace_private *ap = netdev_priv(dev);
2941 struct ace_regs __iomem *regs = ap->regs;
2943 if (!(readl(®s->CpuCtrl) & CPU_HALTED)) {
2944 printk(KERN_ERR "%s: trying to download firmware while the "
2945 "CPU is running!\n", ap->name);
2950 * Do not try to clear more than 512KB or we end up seeing
2951 * funny things on NICs with only 512KB SRAM
2953 ace_clear(regs, 0x2000, 0x80000-0x2000);
2954 if (ACE_IS_TIGON_I(ap)) {
2955 ace_copy(regs, tigonFwText, tigonFwTextAddr, tigonFwTextLen);
2956 ace_copy(regs, tigonFwData, tigonFwDataAddr, tigonFwDataLen);
2957 ace_copy(regs, tigonFwRodata, tigonFwRodataAddr,
2959 ace_clear(regs, tigonFwBssAddr, tigonFwBssLen);
2960 ace_clear(regs, tigonFwSbssAddr, tigonFwSbssLen);
2961 }else if (ap->version == 2) {
2962 ace_clear(regs, tigon2FwBssAddr, tigon2FwBssLen);
2963 ace_clear(regs, tigon2FwSbssAddr, tigon2FwSbssLen);
2964 ace_copy(regs, tigon2FwText, tigon2FwTextAddr,tigon2FwTextLen);
2965 ace_copy(regs, tigon2FwRodata, tigon2FwRodataAddr,
2967 ace_copy(regs, tigon2FwData, tigon2FwDataAddr,tigon2FwDataLen);
2975 * The eeprom on the AceNIC is an Atmel i2c EEPROM.
2977 * Accessing the EEPROM is `interesting' to say the least - don't read
2978 * this code right after dinner.
2980 * This is all about black magic and bit-banging the device .... I
2981 * wonder in what hospital they have put the guy who designed the i2c
2984 * Oh yes, this is only the beginning!
2986 * Thanks to Stevarino Webinski for helping tracking down the bugs in the
2987 * code i2c readout code by beta testing all my hacks.
2989 static void __devinit eeprom_start(struct ace_regs __iomem *regs)
2993 readl(®s->LocalCtrl);
2994 udelay(ACE_SHORT_DELAY);
2995 local = readl(®s->LocalCtrl);
2996 local |= EEPROM_DATA_OUT | EEPROM_WRITE_ENABLE;
2997 writel(local, ®s->LocalCtrl);
2998 readl(®s->LocalCtrl);
3000 udelay(ACE_SHORT_DELAY);
3001 local |= EEPROM_CLK_OUT;
3002 writel(local, ®s->LocalCtrl);
3003 readl(®s->LocalCtrl);
3005 udelay(ACE_SHORT_DELAY);
3006 local &= ~EEPROM_DATA_OUT;
3007 writel(local, ®s->LocalCtrl);
3008 readl(®s->LocalCtrl);
3010 udelay(ACE_SHORT_DELAY);
3011 local &= ~EEPROM_CLK_OUT;
3012 writel(local, ®s->LocalCtrl);
3013 readl(®s->LocalCtrl);
3018 static void __devinit eeprom_prep(struct ace_regs __iomem *regs, u8 magic)
3023 udelay(ACE_SHORT_DELAY);
3024 local = readl(®s->LocalCtrl);
3025 local &= ~EEPROM_DATA_OUT;
3026 local |= EEPROM_WRITE_ENABLE;
3027 writel(local, ®s->LocalCtrl);
3028 readl(®s->LocalCtrl);
3031 for (i = 0; i < 8; i++, magic <<= 1) {
3032 udelay(ACE_SHORT_DELAY);
3034 local |= EEPROM_DATA_OUT;
3036 local &= ~EEPROM_DATA_OUT;
3037 writel(local, ®s->LocalCtrl);
3038 readl(®s->LocalCtrl);
3041 udelay(ACE_SHORT_DELAY);
3042 local |= EEPROM_CLK_OUT;
3043 writel(local, ®s->LocalCtrl);
3044 readl(®s->LocalCtrl);
3046 udelay(ACE_SHORT_DELAY);
3047 local &= ~(EEPROM_CLK_OUT | EEPROM_DATA_OUT);
3048 writel(local, ®s->LocalCtrl);
3049 readl(®s->LocalCtrl);
3055 static int __devinit eeprom_check_ack(struct ace_regs __iomem *regs)
3060 local = readl(®s->LocalCtrl);
3061 local &= ~EEPROM_WRITE_ENABLE;
3062 writel(local, ®s->LocalCtrl);
3063 readl(®s->LocalCtrl);
3065 udelay(ACE_LONG_DELAY);
3066 local |= EEPROM_CLK_OUT;
3067 writel(local, ®s->LocalCtrl);
3068 readl(®s->LocalCtrl);
3070 udelay(ACE_SHORT_DELAY);
3071 /* sample data in middle of high clk */
3072 state = (readl(®s->LocalCtrl) & EEPROM_DATA_IN) != 0;
3073 udelay(ACE_SHORT_DELAY);
3075 writel(readl(®s->LocalCtrl) & ~EEPROM_CLK_OUT, ®s->LocalCtrl);
3076 readl(®s->LocalCtrl);
3083 static void __devinit eeprom_stop(struct ace_regs __iomem *regs)
3087 udelay(ACE_SHORT_DELAY);
3088 local = readl(®s->LocalCtrl);
3089 local |= EEPROM_WRITE_ENABLE;
3090 writel(local, ®s->LocalCtrl);
3091 readl(®s->LocalCtrl);
3093 udelay(ACE_SHORT_DELAY);
3094 local &= ~EEPROM_DATA_OUT;
3095 writel(local, ®s->LocalCtrl);
3096 readl(®s->LocalCtrl);
3098 udelay(ACE_SHORT_DELAY);
3099 local |= EEPROM_CLK_OUT;
3100 writel(local, ®s->LocalCtrl);
3101 readl(®s->LocalCtrl);
3103 udelay(ACE_SHORT_DELAY);
3104 local |= EEPROM_DATA_OUT;
3105 writel(local, ®s->LocalCtrl);
3106 readl(®s->LocalCtrl);
3108 udelay(ACE_LONG_DELAY);
3109 local &= ~EEPROM_CLK_OUT;
3110 writel(local, ®s->LocalCtrl);
3116 * Read a whole byte from the EEPROM.
3118 static int __devinit read_eeprom_byte(struct net_device *dev,
3119 unsigned long offset)
3121 struct ace_private *ap = netdev_priv(dev);
3122 struct ace_regs __iomem *regs = ap->regs;
3123 unsigned long flags;
3129 * Don't take interrupts on this CPU will bit banging
3130 * the %#%#@$ I2C device
3132 local_irq_save(flags);
3136 eeprom_prep(regs, EEPROM_WRITE_SELECT);
3137 if (eeprom_check_ack(regs)) {
3138 local_irq_restore(flags);
3139 printk(KERN_ERR "%s: Unable to sync eeprom\n", ap->name);
3141 goto eeprom_read_error;
3144 eeprom_prep(regs, (offset >> 8) & 0xff);
3145 if (eeprom_check_ack(regs)) {
3146 local_irq_restore(flags);
3147 printk(KERN_ERR "%s: Unable to set address byte 0\n",
3150 goto eeprom_read_error;
3153 eeprom_prep(regs, offset & 0xff);
3154 if (eeprom_check_ack(regs)) {
3155 local_irq_restore(flags);
3156 printk(KERN_ERR "%s: Unable to set address byte 1\n",
3159 goto eeprom_read_error;
3163 eeprom_prep(regs, EEPROM_READ_SELECT);
3164 if (eeprom_check_ack(regs)) {
3165 local_irq_restore(flags);
3166 printk(KERN_ERR "%s: Unable to set READ_SELECT\n",
3169 goto eeprom_read_error;
3172 for (i = 0; i < 8; i++) {
3173 local = readl(®s->LocalCtrl);
3174 local &= ~EEPROM_WRITE_ENABLE;
3175 writel(local, ®s->LocalCtrl);
3176 readl(®s->LocalCtrl);
3177 udelay(ACE_LONG_DELAY);
3179 local |= EEPROM_CLK_OUT;
3180 writel(local, ®s->LocalCtrl);
3181 readl(®s->LocalCtrl);
3183 udelay(ACE_SHORT_DELAY);
3184 /* sample data mid high clk */
3185 result = (result << 1) |
3186 ((readl(®s->LocalCtrl) & EEPROM_DATA_IN) != 0);
3187 udelay(ACE_SHORT_DELAY);
3189 local = readl(®s->LocalCtrl);
3190 local &= ~EEPROM_CLK_OUT;
3191 writel(local, ®s->LocalCtrl);
3192 readl(®s->LocalCtrl);
3193 udelay(ACE_SHORT_DELAY);
3196 local |= EEPROM_WRITE_ENABLE;
3197 writel(local, ®s->LocalCtrl);
3198 readl(®s->LocalCtrl);
3200 udelay(ACE_SHORT_DELAY);
3204 local |= EEPROM_DATA_OUT;
3205 writel(local, ®s->LocalCtrl);
3206 readl(®s->LocalCtrl);
3208 udelay(ACE_SHORT_DELAY);
3209 writel(readl(®s->LocalCtrl) | EEPROM_CLK_OUT, ®s->LocalCtrl);
3210 readl(®s->LocalCtrl);
3211 udelay(ACE_LONG_DELAY);
3212 writel(readl(®s->LocalCtrl) & ~EEPROM_CLK_OUT, ®s->LocalCtrl);
3213 readl(®s->LocalCtrl);
3215 udelay(ACE_SHORT_DELAY);
3218 local_irq_restore(flags);
3223 printk(KERN_ERR "%s: Unable to read eeprom byte 0x%02lx\n",