1/*
2 * acenic.c: Linux driver for the Alteon AceNIC Gigabit Ethernet card
3 *           and other Tigon based cards.
4 *
5 * Copyright 1998-2002 by Jes Sorensen, <jes@trained-monkey.org>.
6 *
7 * Thanks to Alteon and 3Com for providing hardware and documentation
8 * enabling me to write this driver.
9 *
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.
14 *
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.
19 *
20 * Additional credits:
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
44 *                                       endian systems.
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
49 *                                       driver init path.
50 *   Grant Grundler <grundler@cup.hp.com>: PCI write posting fixes.
51 */
52
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/delay.h>
65#include <linux/mm.h>
66#include <linux/highmem.h>
67#include <linux/sockios.h>
68#include <linux/firmware.h>
69#include <linux/slab.h>
70#include <linux/prefetch.h>
71#include <linux/if_vlan.h>
72
73#ifdef SIOCETHTOOL
74#include <linux/ethtool.h>
75#endif
76
77#include <net/sock.h>
78#include <net/ip.h>
79
80#include <asm/io.h>
81#include <asm/irq.h>
82#include <asm/byteorder.h>
83#include <asm/uaccess.h>
84
85
86#define DRV_NAME "acenic"
87
88#undef INDEX_DEBUG
89
90#ifdef CONFIG_ACENIC_OMIT_TIGON_I
91#define ACE_IS_TIGON_I(ap)	0
92#define ACE_TX_RING_ENTRIES(ap)	MAX_TX_RING_ENTRIES
93#else
94#define ACE_IS_TIGON_I(ap)	(ap->version == 1)
95#define ACE_TX_RING_ENTRIES(ap)	ap->tx_ring_entries
96#endif
97
98#ifndef PCI_VENDOR_ID_ALTEON
99#define PCI_VENDOR_ID_ALTEON		0x12ae
100#endif
101#ifndef PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE
102#define PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE  0x0001
103#define PCI_DEVICE_ID_ALTEON_ACENIC_COPPER 0x0002
104#endif
105#ifndef PCI_DEVICE_ID_3COM_3C985
106#define PCI_DEVICE_ID_3COM_3C985	0x0001
107#endif
108#ifndef PCI_VENDOR_ID_NETGEAR
109#define PCI_VENDOR_ID_NETGEAR		0x1385
110#define PCI_DEVICE_ID_NETGEAR_GA620	0x620a
111#endif
112#ifndef PCI_DEVICE_ID_NETGEAR_GA620T
113#define PCI_DEVICE_ID_NETGEAR_GA620T	0x630a
114#endif
115
116
117/*
118 * Farallon used the DEC vendor ID by mistake and they seem not
119 * to care - stinky!
120 */
121#ifndef PCI_DEVICE_ID_FARALLON_PN9000SX
122#define PCI_DEVICE_ID_FARALLON_PN9000SX	0x1a
123#endif
124#ifndef PCI_DEVICE_ID_FARALLON_PN9100T
125#define PCI_DEVICE_ID_FARALLON_PN9100T  0xfa
126#endif
127#ifndef PCI_VENDOR_ID_SGI
128#define PCI_VENDOR_ID_SGI		0x10a9
129#endif
130#ifndef PCI_DEVICE_ID_SGI_ACENIC
131#define PCI_DEVICE_ID_SGI_ACENIC	0x0009
132#endif
133
134static const struct pci_device_id acenic_pci_tbl[] = {
135	{ PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE,
136	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
137	{ PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_ALTEON_ACENIC_COPPER,
138	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
139	{ PCI_VENDOR_ID_3COM, PCI_DEVICE_ID_3COM_3C985,
140	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
141	{ PCI_VENDOR_ID_NETGEAR, PCI_DEVICE_ID_NETGEAR_GA620,
142	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
143	{ PCI_VENDOR_ID_NETGEAR, PCI_DEVICE_ID_NETGEAR_GA620T,
144	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
145	/*
146	 * Farallon used the DEC vendor ID on their cards incorrectly,
147	 * then later Alteon's ID.
148	 */
149	{ PCI_VENDOR_ID_DEC, PCI_DEVICE_ID_FARALLON_PN9000SX,
150	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
151	{ PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_FARALLON_PN9100T,
152	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
153	{ PCI_VENDOR_ID_SGI, PCI_DEVICE_ID_SGI_ACENIC,
154	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
155	{ }
156};
157MODULE_DEVICE_TABLE(pci, acenic_pci_tbl);
158
159#define ace_sync_irq(irq)	synchronize_irq(irq)
160
161#ifndef offset_in_page
162#define offset_in_page(ptr)	((unsigned long)(ptr) & ~PAGE_MASK)
163#endif
164
165#define ACE_MAX_MOD_PARMS	8
166#define BOARD_IDX_STATIC	0
167#define BOARD_IDX_OVERFLOW	-1
168
169#include "acenic.h"
170
171/*
172 * These must be defined before the firmware is included.
173 */
174#define MAX_TEXT_LEN	96*1024
175#define MAX_RODATA_LEN	8*1024
176#define MAX_DATA_LEN	2*1024
177
178#ifndef tigon2FwReleaseLocal
179#define tigon2FwReleaseLocal 0
180#endif
181
182/*
183 * This driver currently supports Tigon I and Tigon II based cards
184 * including the Alteon AceNIC, the 3Com 3C985[B] and NetGear
185 * GA620. The driver should also work on the SGI, DEC and Farallon
186 * versions of the card, however I have not been able to test that
187 * myself.
188 *
189 * This card is really neat, it supports receive hardware checksumming
190 * and jumbo frames (up to 9000 bytes) and does a lot of work in the
191 * firmware. Also the programming interface is quite neat, except for
192 * the parts dealing with the i2c eeprom on the card ;-)
193 *
194 * Using jumbo frames:
195 *
196 * To enable jumbo frames, simply specify an mtu between 1500 and 9000
197 * bytes to ifconfig. Jumbo frames can be enabled or disabled at any time
198 * by running `ifconfig eth<X> mtu <MTU>' with <X> being the Ethernet
199 * interface number and <MTU> being the MTU value.
200 *
201 * Module parameters:
202 *
203 * When compiled as a loadable module, the driver allows for a number
204 * of module parameters to be specified. The driver supports the
205 * following module parameters:
206 *
207 *  trace=<val> - Firmware trace level. This requires special traced
208 *                firmware to replace the firmware supplied with
209 *                the driver - for debugging purposes only.
210 *
211 *  link=<val>  - Link state. Normally you want to use the default link
212 *                parameters set by the driver. This can be used to
213 *                override these in case your switch doesn't negotiate
214 *                the link properly. Valid values are:
215 *         0x0001 - Force half duplex link.
216 *         0x0002 - Do not negotiate line speed with the other end.
217 *         0x0010 - 10Mbit/sec link.
218 *         0x0020 - 100Mbit/sec link.
219 *         0x0040 - 1000Mbit/sec link.
220 *         0x0100 - Do not negotiate flow control.
221 *         0x0200 - Enable RX flow control Y
222 *         0x0400 - Enable TX flow control Y (Tigon II NICs only).
223 *                Default value is 0x0270, ie. enable link+flow
224 *                control negotiation. Negotiating the highest
225 *                possible link speed with RX flow control enabled.
226 *
227 *                When disabling link speed negotiation, only one link
228 *                speed is allowed to be specified!
229 *
230 *  tx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
231 *                to wait for more packets to arive before
232 *                interrupting the host, from the time the first
233 *                packet arrives.
234 *
235 *  rx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
236 *                to wait for more packets to arive in the transmit ring,
237 *                before interrupting the host, after transmitting the
238 *                first packet in the ring.
239 *
240 *  max_tx_desc=<val> - maximum number of transmit descriptors
241 *                (packets) transmitted before interrupting the host.
242 *
243 *  max_rx_desc=<val> - maximum number of receive descriptors
244 *                (packets) received before interrupting the host.
245 *
246 *  tx_ratio=<val> - 7 bit value (0 - 63) specifying the split in 64th
247 *                increments of the NIC's on board memory to be used for
248 *                transmit and receive buffers. For the 1MB NIC app. 800KB
249 *                is available, on the 1/2MB NIC app. 300KB is available.
250 *                68KB will always be available as a minimum for both
251 *                directions. The default value is a 50/50 split.
252 *  dis_pci_mem_inval=<val> - disable PCI memory write and invalidate
253 *                operations, default (1) is to always disable this as
254 *                that is what Alteon does on NT. I have not been able
255 *                to measure any real performance differences with
256 *                this on my systems. Set <val>=0 if you want to
257 *                enable these operations.
258 *
259 * If you use more than one NIC, specify the parameters for the
260 * individual NICs with a comma, ie. trace=0,0x00001fff,0 you want to
261 * run tracing on NIC #2 but not on NIC #1 and #3.
262 *
263 * TODO:
264 *
265 * - Proper multicast support.
266 * - NIC dump support.
267 * - More tuning parameters.
268 *
269 * The mini ring is not used under Linux and I am not sure it makes sense
270 * to actually use it.
271 *
272 * New interrupt handler strategy:
273 *
274 * The old interrupt handler worked using the traditional method of
275 * replacing an skbuff with a new one when a packet arrives. However
276 * the rx rings do not need to contain a static number of buffer
277 * descriptors, thus it makes sense to move the memory allocation out
278 * of the main interrupt handler and do it in a bottom half handler
279 * and only allocate new buffers when the number of buffers in the
280 * ring is below a certain threshold. In order to avoid starving the
281 * NIC under heavy load it is however necessary to force allocation
282 * when hitting a minimum threshold. The strategy for alloction is as
283 * follows:
284 *
285 *     RX_LOW_BUF_THRES    - allocate buffers in the bottom half
286 *     RX_PANIC_LOW_THRES  - we are very low on buffers, allocate
287 *                           the buffers in the interrupt handler
288 *     RX_RING_THRES       - maximum number of buffers in the rx ring
289 *     RX_MINI_THRES       - maximum number of buffers in the mini ring
290 *     RX_JUMBO_THRES      - maximum number of buffers in the jumbo ring
291 *
292 * One advantagous side effect of this allocation approach is that the
293 * entire rx processing can be done without holding any spin lock
294 * since the rx rings and registers are totally independent of the tx
295 * ring and its registers.  This of course includes the kmalloc's of
296 * new skb's. Thus start_xmit can run in parallel with rx processing
297 * and the memory allocation on SMP systems.
298 *
299 * Note that running the skb reallocation in a bottom half opens up
300 * another can of races which needs to be handled properly. In
301 * particular it can happen that the interrupt handler tries to run
302 * the reallocation while the bottom half is either running on another
303 * CPU or was interrupted on the same CPU. To get around this the
304 * driver uses bitops to prevent the reallocation routines from being
305 * reentered.
306 *
307 * TX handling can also be done without holding any spin lock, wheee
308 * this is fun! since tx_ret_csm is only written to by the interrupt
309 * handler. The case to be aware of is when shutting down the device
310 * and cleaning up where it is necessary to make sure that
311 * start_xmit() is not running while this is happening. Well DaveM
312 * informs me that this case is already protected against ... bye bye
313 * Mr. Spin Lock, it was nice to know you.
314 *
315 * TX interrupts are now partly disabled so the NIC will only generate
316 * TX interrupts for the number of coal ticks, not for the number of
317 * TX packets in the queue. This should reduce the number of TX only,
318 * ie. when no RX processing is done, interrupts seen.
319 */
320
321/*
322 * Threshold values for RX buffer allocation - the low water marks for
323 * when to start refilling the rings are set to 75% of the ring
324 * sizes. It seems to make sense to refill the rings entirely from the
325 * intrrupt handler once it gets below the panic threshold, that way
326 * we don't risk that the refilling is moved to another CPU when the
327 * one running the interrupt handler just got the slab code hot in its
328 * cache.
329 */
330#define RX_RING_SIZE		72
331#define RX_MINI_SIZE		64
332#define RX_JUMBO_SIZE		48
333
334#define RX_PANIC_STD_THRES	16
335#define RX_PANIC_STD_REFILL	(3*RX_PANIC_STD_THRES)/2
336#define RX_LOW_STD_THRES	(3*RX_RING_SIZE)/4
337#define RX_PANIC_MINI_THRES	12
338#define RX_PANIC_MINI_REFILL	(3*RX_PANIC_MINI_THRES)/2
339#define RX_LOW_MINI_THRES	(3*RX_MINI_SIZE)/4
340#define RX_PANIC_JUMBO_THRES	6
341#define RX_PANIC_JUMBO_REFILL	(3*RX_PANIC_JUMBO_THRES)/2
342#define RX_LOW_JUMBO_THRES	(3*RX_JUMBO_SIZE)/4
343
344
345/*
346 * Size of the mini ring entries, basically these just should be big
347 * enough to take TCP ACKs
348 */
349#define ACE_MINI_SIZE		100
350
351#define ACE_MINI_BUFSIZE	ACE_MINI_SIZE
352#define ACE_STD_BUFSIZE		(ACE_STD_MTU + ETH_HLEN + 4)
353#define ACE_JUMBO_BUFSIZE	(ACE_JUMBO_MTU + ETH_HLEN + 4)
354
355/*
356 * There seems to be a magic difference in the effect between 995 and 996
357 * but little difference between 900 and 995 ... no idea why.
358 *
359 * There is now a default set of tuning parameters which is set, depending
360 * on whether or not the user enables Jumbo frames. It's assumed that if
361 * Jumbo frames are enabled, the user wants optimal tuning for that case.
362 */
363#define DEF_TX_COAL		400 /* 996 */
364#define DEF_TX_MAX_DESC		60  /* was 40 */
365#define DEF_RX_COAL		120 /* 1000 */
366#define DEF_RX_MAX_DESC		25
367#define DEF_TX_RATIO		21 /* 24 */
368
369#define DEF_JUMBO_TX_COAL	20
370#define DEF_JUMBO_TX_MAX_DESC	60
371#define DEF_JUMBO_RX_COAL	30
372#define DEF_JUMBO_RX_MAX_DESC	6
373#define DEF_JUMBO_TX_RATIO	21
374
375#if tigon2FwReleaseLocal < 20001118
376/*
377 * Standard firmware and early modifications duplicate
378 * IRQ load without this flag (coal timer is never reset).
379 * Note that with this flag tx_coal should be less than
380 * time to xmit full tx ring.
381 * 400usec is not so bad for tx ring size of 128.
382 */
383#define TX_COAL_INTS_ONLY	1	/* worth it */
384#else
385/*
386 * With modified firmware, this is not necessary, but still useful.
387 */
388#define TX_COAL_INTS_ONLY	1
389#endif
390
391#define DEF_TRACE		0
392#define DEF_STAT		(2 * TICKS_PER_SEC)
393
394
395static int link_state[ACE_MAX_MOD_PARMS];
396static int trace[ACE_MAX_MOD_PARMS];
397static int tx_coal_tick[ACE_MAX_MOD_PARMS];
398static int rx_coal_tick[ACE_MAX_MOD_PARMS];
399static int max_tx_desc[ACE_MAX_MOD_PARMS];
400static int max_rx_desc[ACE_MAX_MOD_PARMS];
401static int tx_ratio[ACE_MAX_MOD_PARMS];
402static int dis_pci_mem_inval[ACE_MAX_MOD_PARMS] = {1, 1, 1, 1, 1, 1, 1, 1};
403
404MODULE_AUTHOR("Jes Sorensen <jes@trained-monkey.org>");
405MODULE_LICENSE("GPL");
406MODULE_DESCRIPTION("AceNIC/3C985/GA620 Gigabit Ethernet driver");
407#ifndef CONFIG_ACENIC_OMIT_TIGON_I
408MODULE_FIRMWARE("acenic/tg1.bin");
409#endif
410MODULE_FIRMWARE("acenic/tg2.bin");
411
412module_param_array_named(link, link_state, int, NULL, 0);
413module_param_array(trace, int, NULL, 0);
414module_param_array(tx_coal_tick, int, NULL, 0);
415module_param_array(max_tx_desc, int, NULL, 0);
416module_param_array(rx_coal_tick, int, NULL, 0);
417module_param_array(max_rx_desc, int, NULL, 0);
418module_param_array(tx_ratio, int, NULL, 0);
419MODULE_PARM_DESC(link, "AceNIC/3C985/NetGear link state");
420MODULE_PARM_DESC(trace, "AceNIC/3C985/NetGear firmware trace level");
421MODULE_PARM_DESC(tx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first tx descriptor arrives");
422MODULE_PARM_DESC(max_tx_desc, "AceNIC/3C985/GA620 max number of transmit descriptors to wait");
423MODULE_PARM_DESC(rx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first rx descriptor arrives");
424MODULE_PARM_DESC(max_rx_desc, "AceNIC/3C985/GA620 max number of receive descriptors to wait");
425MODULE_PARM_DESC(tx_ratio, "AceNIC/3C985/GA620 ratio of NIC memory used for TX/RX descriptors (range 0-63)");
426
427
428static const char version[] =
429  "acenic.c: v0.92 08/05/2002  Jes Sorensen, linux-acenic@SunSITE.dk\n"
430  "                            http://home.cern.ch/~jes/gige/acenic.html\n";
431
432static int ace_get_settings(struct net_device *, struct ethtool_cmd *);
433static int ace_set_settings(struct net_device *, struct ethtool_cmd *);
434static void ace_get_drvinfo(struct net_device *, struct ethtool_drvinfo *);
435
436static const struct ethtool_ops ace_ethtool_ops = {
437	.get_settings = ace_get_settings,
438	.set_settings = ace_set_settings,
439	.get_drvinfo = ace_get_drvinfo,
440};
441
442static void ace_watchdog(struct net_device *dev);
443
444static const struct net_device_ops ace_netdev_ops = {
445	.ndo_open		= ace_open,
446	.ndo_stop		= ace_close,
447	.ndo_tx_timeout		= ace_watchdog,
448	.ndo_get_stats		= ace_get_stats,
449	.ndo_start_xmit		= ace_start_xmit,
450	.ndo_set_rx_mode	= ace_set_multicast_list,
451	.ndo_validate_addr	= eth_validate_addr,
452	.ndo_set_mac_address	= ace_set_mac_addr,
453	.ndo_change_mtu		= ace_change_mtu,
454};
455
456static int acenic_probe_one(struct pci_dev *pdev,
457			    const struct pci_device_id *id)
458{
459	struct net_device *dev;
460	struct ace_private *ap;
461	static int boards_found;
462
463	dev = alloc_etherdev(sizeof(struct ace_private));
464	if (dev == NULL)
465		return -ENOMEM;
466
467	SET_NETDEV_DEV(dev, &pdev->dev);
468
469	ap = netdev_priv(dev);
470	ap->pdev = pdev;
471	ap->name = pci_name(pdev);
472
473	dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
474	dev->features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX;
475
476	dev->watchdog_timeo = 5*HZ;
477
478	dev->netdev_ops = &ace_netdev_ops;
479	dev->ethtool_ops = &ace_ethtool_ops;
480
481	/* we only display this string ONCE */
482	if (!boards_found)
483		printk(version);
484
485	if (pci_enable_device(pdev))
486		goto fail_free_netdev;
487
488	/*
489	 * Enable master mode before we start playing with the
490	 * pci_command word since pci_set_master() will modify
491	 * it.
492	 */
493	pci_set_master(pdev);
494
495	pci_read_config_word(pdev, PCI_COMMAND, &ap->pci_command);
496
497	/* OpenFirmware on Mac's does not set this - DOH.. */
498	if (!(ap->pci_command & PCI_COMMAND_MEMORY)) {
499		printk(KERN_INFO "%s: Enabling PCI Memory Mapped "
500		       "access - was not enabled by BIOS/Firmware\n",
501		       ap->name);
502		ap->pci_command = ap->pci_command | PCI_COMMAND_MEMORY;
503		pci_write_config_word(ap->pdev, PCI_COMMAND,
504				      ap->pci_command);
505		wmb();
506	}
507
508	pci_read_config_byte(pdev, PCI_LATENCY_TIMER, &ap->pci_latency);
509	if (ap->pci_latency <= 0x40) {
510		ap->pci_latency = 0x40;
511		pci_write_config_byte(pdev, PCI_LATENCY_TIMER, ap->pci_latency);
512	}
513
514	/*
515	 * Remap the regs into kernel space - this is abuse of
516	 * dev->base_addr since it was means for I/O port
517	 * addresses but who gives a damn.
518	 */
519	dev->base_addr = pci_resource_start(pdev, 0);
520	ap->regs = ioremap(dev->base_addr, 0x4000);
521	if (!ap->regs) {
522		printk(KERN_ERR "%s:  Unable to map I/O register, "
523		       "AceNIC %i will be disabled.\n",
524		       ap->name, boards_found);
525		goto fail_free_netdev;
526	}
527
528	switch(pdev->vendor) {
529	case PCI_VENDOR_ID_ALTEON:
530		if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9100T) {
531			printk(KERN_INFO "%s: Farallon PN9100-T ",
532			       ap->name);
533		} else {
534			printk(KERN_INFO "%s: Alteon AceNIC ",
535			       ap->name);
536		}
537		break;
538	case PCI_VENDOR_ID_3COM:
539		printk(KERN_INFO "%s: 3Com 3C985 ", ap->name);
540		break;
541	case PCI_VENDOR_ID_NETGEAR:
542		printk(KERN_INFO "%s: NetGear GA620 ", ap->name);
543		break;
544	case PCI_VENDOR_ID_DEC:
545		if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9000SX) {
546			printk(KERN_INFO "%s: Farallon PN9000-SX ",
547			       ap->name);
548			break;
549		}
550	case PCI_VENDOR_ID_SGI:
551		printk(KERN_INFO "%s: SGI AceNIC ", ap->name);
552		break;
553	default:
554		printk(KERN_INFO "%s: Unknown AceNIC ", ap->name);
555		break;
556	}
557
558	printk("Gigabit Ethernet at 0x%08lx, ", dev->base_addr);
559	printk("irq %d\n", pdev->irq);
560
561#ifdef CONFIG_ACENIC_OMIT_TIGON_I
562	if ((readl(&ap->regs->HostCtrl) >> 28) == 4) {
563		printk(KERN_ERR "%s: Driver compiled without Tigon I"
564		       " support - NIC disabled\n", dev->name);
565		goto fail_uninit;
566	}
567#endif
568
569	if (ace_allocate_descriptors(dev))
570		goto fail_free_netdev;
571
572#ifdef MODULE
573	if (boards_found >= ACE_MAX_MOD_PARMS)
574		ap->board_idx = BOARD_IDX_OVERFLOW;
575	else
576		ap->board_idx = boards_found;
577#else
578	ap->board_idx = BOARD_IDX_STATIC;
579#endif
580
581	if (ace_init(dev))
582		goto fail_free_netdev;
583
584	if (register_netdev(dev)) {
585		printk(KERN_ERR "acenic: device registration failed\n");
586		goto fail_uninit;
587	}
588	ap->name = dev->name;
589
590	if (ap->pci_using_dac)
591		dev->features |= NETIF_F_HIGHDMA;
592
593	pci_set_drvdata(pdev, dev);
594
595	boards_found++;
596	return 0;
597
598 fail_uninit:
599	ace_init_cleanup(dev);
600 fail_free_netdev:
601	free_netdev(dev);
602	return -ENODEV;
603}
604
605static void acenic_remove_one(struct pci_dev *pdev)
606{
607	struct net_device *dev = pci_get_drvdata(pdev);
608	struct ace_private *ap = netdev_priv(dev);
609	struct ace_regs __iomem *regs = ap->regs;
610	short i;
611
612	unregister_netdev(dev);
613
614	writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
615	if (ap->version >= 2)
616		writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
617
618	/*
619	 * This clears any pending interrupts
620	 */
621	writel(1, &regs->Mb0Lo);
622	readl(&regs->CpuCtrl);	/* flush */
623
624	/*
625	 * Make sure no other CPUs are processing interrupts
626	 * on the card before the buffers are being released.
627	 * Otherwise one might experience some `interesting'
628	 * effects.
629	 *
630	 * Then release the RX buffers - jumbo buffers were
631	 * already released in ace_close().
632	 */
633	ace_sync_irq(dev->irq);
634
635	for (i = 0; i < RX_STD_RING_ENTRIES; i++) {
636		struct sk_buff *skb = ap->skb->rx_std_skbuff[i].skb;
637
638		if (skb) {
639			struct ring_info *ringp;
640			dma_addr_t mapping;
641
642			ringp = &ap->skb->rx_std_skbuff[i];
643			mapping = dma_unmap_addr(ringp, mapping);
644			pci_unmap_page(ap->pdev, mapping,
645				       ACE_STD_BUFSIZE,
646				       PCI_DMA_FROMDEVICE);
647
648			ap->rx_std_ring[i].size = 0;
649			ap->skb->rx_std_skbuff[i].skb = NULL;
650			dev_kfree_skb(skb);
651		}
652	}
653
654	if (ap->version >= 2) {
655		for (i = 0; i < RX_MINI_RING_ENTRIES; i++) {
656			struct sk_buff *skb = ap->skb->rx_mini_skbuff[i].skb;
657
658			if (skb) {
659				struct ring_info *ringp;
660				dma_addr_t mapping;
661
662				ringp = &ap->skb->rx_mini_skbuff[i];
663				mapping = dma_unmap_addr(ringp,mapping);
664				pci_unmap_page(ap->pdev, mapping,
665					       ACE_MINI_BUFSIZE,
666					       PCI_DMA_FROMDEVICE);
667
668				ap->rx_mini_ring[i].size = 0;
669				ap->skb->rx_mini_skbuff[i].skb = NULL;
670				dev_kfree_skb(skb);
671			}
672		}
673	}
674
675	for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
676		struct sk_buff *skb = ap->skb->rx_jumbo_skbuff[i].skb;
677		if (skb) {
678			struct ring_info *ringp;
679			dma_addr_t mapping;
680
681			ringp = &ap->skb->rx_jumbo_skbuff[i];
682			mapping = dma_unmap_addr(ringp, mapping);
683			pci_unmap_page(ap->pdev, mapping,
684				       ACE_JUMBO_BUFSIZE,
685				       PCI_DMA_FROMDEVICE);
686
687			ap->rx_jumbo_ring[i].size = 0;
688			ap->skb->rx_jumbo_skbuff[i].skb = NULL;
689			dev_kfree_skb(skb);
690		}
691	}
692
693	ace_init_cleanup(dev);
694	free_netdev(dev);
695}
696
697static struct pci_driver acenic_pci_driver = {
698	.name		= "acenic",
699	.id_table	= acenic_pci_tbl,
700	.probe		= acenic_probe_one,
701	.remove		= acenic_remove_one,
702};
703
704static void ace_free_descriptors(struct net_device *dev)
705{
706	struct ace_private *ap = netdev_priv(dev);
707	int size;
708
709	if (ap->rx_std_ring != NULL) {
710		size = (sizeof(struct rx_desc) *
711			(RX_STD_RING_ENTRIES +
712			 RX_JUMBO_RING_ENTRIES +
713			 RX_MINI_RING_ENTRIES +
714			 RX_RETURN_RING_ENTRIES));
715		pci_free_consistent(ap->pdev, size, ap->rx_std_ring,
716				    ap->rx_ring_base_dma);
717		ap->rx_std_ring = NULL;
718		ap->rx_jumbo_ring = NULL;
719		ap->rx_mini_ring = NULL;
720		ap->rx_return_ring = NULL;
721	}
722	if (ap->evt_ring != NULL) {
723		size = (sizeof(struct event) * EVT_RING_ENTRIES);
724		pci_free_consistent(ap->pdev, size, ap->evt_ring,
725				    ap->evt_ring_dma);
726		ap->evt_ring = NULL;
727	}
728	if (ap->tx_ring != NULL && !ACE_IS_TIGON_I(ap)) {
729		size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
730		pci_free_consistent(ap->pdev, size, ap->tx_ring,
731				    ap->tx_ring_dma);
732	}
733	ap->tx_ring = NULL;
734
735	if (ap->evt_prd != NULL) {
736		pci_free_consistent(ap->pdev, sizeof(u32),
737				    (void *)ap->evt_prd, ap->evt_prd_dma);
738		ap->evt_prd = NULL;
739	}
740	if (ap->rx_ret_prd != NULL) {
741		pci_free_consistent(ap->pdev, sizeof(u32),
742				    (void *)ap->rx_ret_prd,
743				    ap->rx_ret_prd_dma);
744		ap->rx_ret_prd = NULL;
745	}
746	if (ap->tx_csm != NULL) {
747		pci_free_consistent(ap->pdev, sizeof(u32),
748				    (void *)ap->tx_csm, ap->tx_csm_dma);
749		ap->tx_csm = NULL;
750	}
751}
752
753
754static int ace_allocate_descriptors(struct net_device *dev)
755{
756	struct ace_private *ap = netdev_priv(dev);
757	int size;
758
759	size = (sizeof(struct rx_desc) *
760		(RX_STD_RING_ENTRIES +
761		 RX_JUMBO_RING_ENTRIES +
762		 RX_MINI_RING_ENTRIES +
763		 RX_RETURN_RING_ENTRIES));
764
765	ap->rx_std_ring = pci_alloc_consistent(ap->pdev, size,
766					       &ap->rx_ring_base_dma);
767	if (ap->rx_std_ring == NULL)
768		goto fail;
769
770	ap->rx_jumbo_ring = ap->rx_std_ring + RX_STD_RING_ENTRIES;
771	ap->rx_mini_ring = ap->rx_jumbo_ring + RX_JUMBO_RING_ENTRIES;
772	ap->rx_return_ring = ap->rx_mini_ring + RX_MINI_RING_ENTRIES;
773
774	size = (sizeof(struct event) * EVT_RING_ENTRIES);
775
776	ap->evt_ring = pci_alloc_consistent(ap->pdev, size, &ap->evt_ring_dma);
777
778	if (ap->evt_ring == NULL)
779		goto fail;
780
781	/*
782	 * Only allocate a host TX ring for the Tigon II, the Tigon I
783	 * has to use PCI registers for this ;-(
784	 */
785	if (!ACE_IS_TIGON_I(ap)) {
786		size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
787
788		ap->tx_ring = pci_alloc_consistent(ap->pdev, size,
789						   &ap->tx_ring_dma);
790
791		if (ap->tx_ring == NULL)
792			goto fail;
793	}
794
795	ap->evt_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
796					   &ap->evt_prd_dma);
797	if (ap->evt_prd == NULL)
798		goto fail;
799
800	ap->rx_ret_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
801					      &ap->rx_ret_prd_dma);
802	if (ap->rx_ret_prd == NULL)
803		goto fail;
804
805	ap->tx_csm = pci_alloc_consistent(ap->pdev, sizeof(u32),
806					  &ap->tx_csm_dma);
807	if (ap->tx_csm == NULL)
808		goto fail;
809
810	return 0;
811
812fail:
813	/* Clean up. */
814	ace_init_cleanup(dev);
815	return 1;
816}
817
818
819/*
820 * Generic cleanup handling data allocated during init. Used when the
821 * module is unloaded or if an error occurs during initialization
822 */
823static void ace_init_cleanup(struct net_device *dev)
824{
825	struct ace_private *ap;
826
827	ap = netdev_priv(dev);
828
829	ace_free_descriptors(dev);
830
831	if (ap->info)
832		pci_free_consistent(ap->pdev, sizeof(struct ace_info),
833				    ap->info, ap->info_dma);
834	kfree(ap->skb);
835	kfree(ap->trace_buf);
836
837	if (dev->irq)
838		free_irq(dev->irq, dev);
839
840	iounmap(ap->regs);
841}
842
843
844/*
845 * Commands are considered to be slow.
846 */
847static inline void ace_issue_cmd(struct ace_regs __iomem *regs, struct cmd *cmd)
848{
849	u32 idx;
850
851	idx = readl(&regs->CmdPrd);
852
853	writel(*(u32 *)(cmd), &regs->CmdRng[idx]);
854	idx = (idx + 1) % CMD_RING_ENTRIES;
855
856	writel(idx, &regs->CmdPrd);
857}
858
859
860static int ace_init(struct net_device *dev)
861{
862	struct ace_private *ap;
863	struct ace_regs __iomem *regs;
864	struct ace_info *info = NULL;
865	struct pci_dev *pdev;
866	unsigned long myjif;
867	u64 tmp_ptr;
868	u32 tig_ver, mac1, mac2, tmp, pci_state;
869	int board_idx, ecode = 0;
870	short i;
871	unsigned char cache_size;
872
873	ap = netdev_priv(dev);
874	regs = ap->regs;
875
876	board_idx = ap->board_idx;
877
878	/*
879	 * aman@sgi.com - its useful to do a NIC reset here to
880	 * address the `Firmware not running' problem subsequent
881	 * to any crashes involving the NIC
882	 */
883	writel(HW_RESET | (HW_RESET << 24), &regs->HostCtrl);
884	readl(&regs->HostCtrl);		/* PCI write posting */
885	udelay(5);
886
887	/*
888	 * Don't access any other registers before this point!
889	 */
890#ifdef __BIG_ENDIAN
891	/*
892	 * This will most likely need BYTE_SWAP once we switch
893	 * to using __raw_writel()
894	 */
895	writel((WORD_SWAP | CLR_INT | ((WORD_SWAP | CLR_INT) << 24)),
896	       &regs->HostCtrl);
897#else
898	writel((CLR_INT | WORD_SWAP | ((CLR_INT | WORD_SWAP) << 24)),
899	       &regs->HostCtrl);
900#endif
901	readl(&regs->HostCtrl);		/* PCI write posting */
902
903	/*
904	 * Stop the NIC CPU and clear pending interrupts
905	 */
906	writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
907	readl(&regs->CpuCtrl);		/* PCI write posting */
908	writel(0, &regs->Mb0Lo);
909
910	tig_ver = readl(&regs->HostCtrl) >> 28;
911
912	switch(tig_ver){
913#ifndef CONFIG_ACENIC_OMIT_TIGON_I
914	case 4:
915	case 5:
916		printk(KERN_INFO "  Tigon I  (Rev. %i), Firmware: %i.%i.%i, ",
917		       tig_ver, ap->firmware_major, ap->firmware_minor,
918		       ap->firmware_fix);
919		writel(0, &regs->LocalCtrl);
920		ap->version = 1;
921		ap->tx_ring_entries = TIGON_I_TX_RING_ENTRIES;
922		break;
923#endif
924	case 6:
925		printk(KERN_INFO "  Tigon II (Rev. %i), Firmware: %i.%i.%i, ",
926		       tig_ver, ap->firmware_major, ap->firmware_minor,
927		       ap->firmware_fix);
928		writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
929		readl(&regs->CpuBCtrl);		/* PCI write posting */
930		/*
931		 * The SRAM bank size does _not_ indicate the amount
932		 * of memory on the card, it controls the _bank_ size!
933		 * Ie. a 1MB AceNIC will have two banks of 512KB.
934		 */
935		writel(SRAM_BANK_512K, &regs->LocalCtrl);
936		writel(SYNC_SRAM_TIMING, &regs->MiscCfg);
937		ap->version = 2;
938		ap->tx_ring_entries = MAX_TX_RING_ENTRIES;
939		break;
940	default:
941		printk(KERN_WARNING "  Unsupported Tigon version detected "
942		       "(%i)\n", tig_ver);
943		ecode = -ENODEV;
944		goto init_error;
945	}
946
947	/*
948	 * ModeStat _must_ be set after the SRAM settings as this change
949	 * seems to corrupt the ModeStat and possible other registers.
950	 * The SRAM settings survive resets and setting it to the same
951	 * value a second time works as well. This is what caused the
952	 * `Firmware not running' problem on the Tigon II.
953	 */
954#ifdef __BIG_ENDIAN
955	writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL | ACE_BYTE_SWAP_BD |
956	       ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
957#else
958	writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL |
959	       ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
960#endif
961	readl(&regs->ModeStat);		/* PCI write posting */
962
963	mac1 = 0;
964	for(i = 0; i < 4; i++) {
965		int t;
966
967		mac1 = mac1 << 8;
968		t = read_eeprom_byte(dev, 0x8c+i);
969		if (t < 0) {
970			ecode = -EIO;
971			goto init_error;
972		} else
973			mac1 |= (t & 0xff);
974	}
975	mac2 = 0;
976	for(i = 4; i < 8; i++) {
977		int t;
978
979		mac2 = mac2 << 8;
980		t = read_eeprom_byte(dev, 0x8c+i);
981		if (t < 0) {
982			ecode = -EIO;
983			goto init_error;
984		} else
985			mac2 |= (t & 0xff);
986	}
987
988	writel(mac1, &regs->MacAddrHi);
989	writel(mac2, &regs->MacAddrLo);
990
991	dev->dev_addr[0] = (mac1 >> 8) & 0xff;
992	dev->dev_addr[1] = mac1 & 0xff;
993	dev->dev_addr[2] = (mac2 >> 24) & 0xff;
994	dev->dev_addr[3] = (mac2 >> 16) & 0xff;
995	dev->dev_addr[4] = (mac2 >> 8) & 0xff;
996	dev->dev_addr[5] = mac2 & 0xff;
997
998	printk("MAC: %pM\n", dev->dev_addr);
999
1000	/*
1001	 * Looks like this is necessary to deal with on all architectures,
1002	 * even this %$#%$# N440BX Intel based thing doesn't get it right.
1003	 * Ie. having two NICs in the machine, one will have the cache
1004	 * line set at boot time, the other will not.
1005	 */
1006	pdev = ap->pdev;
1007	pci_read_config_byte(pdev, PCI_CACHE_LINE_SIZE, &cache_size);
1008	cache_size <<= 2;
1009	if (cache_size != SMP_CACHE_BYTES) {
1010		printk(KERN_INFO "  PCI cache line size set incorrectly "
1011		       "(%i bytes) by BIOS/FW, ", cache_size);
1012		if (cache_size > SMP_CACHE_BYTES)
1013			printk("expecting %i\n", SMP_CACHE_BYTES);
1014		else {
1015			printk("correcting to %i\n", SMP_CACHE_BYTES);
1016			pci_write_config_byte(pdev, PCI_CACHE_LINE_SIZE,
1017					      SMP_CACHE_BYTES >> 2);
1018		}
1019	}
1020
1021	pci_state = readl(&regs->PciState);
1022	printk(KERN_INFO "  PCI bus width: %i bits, speed: %iMHz, "
1023	       "latency: %i clks\n",
1024	       	(pci_state & PCI_32BIT) ? 32 : 64,
1025		(pci_state & PCI_66MHZ) ? 66 : 33,
1026		ap->pci_latency);
1027
1028	/*
1029	 * Set the max DMA transfer size. Seems that for most systems
1030	 * the performance is better when no MAX parameter is
1031	 * set. However for systems enabling PCI write and invalidate,
1032	 * DMA writes must be set to the L1 cache line size to get
1033	 * optimal performance.
1034	 *
1035	 * The default is now to turn the PCI write and invalidate off
1036	 * - that is what Alteon does for NT.
1037	 */
1038	tmp = READ_CMD_MEM | WRITE_CMD_MEM;
1039	if (ap->version >= 2) {
1040		tmp |= (MEM_READ_MULTIPLE | (pci_state & PCI_66MHZ));
1041		/*
1042		 * Tuning parameters only supported for 8 cards
1043		 */
1044		if (board_idx == BOARD_IDX_OVERFLOW ||
1045		    dis_pci_mem_inval[board_idx]) {
1046			if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1047				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1048				pci_write_config_word(pdev, PCI_COMMAND,
1049						      ap->pci_command);
1050				printk(KERN_INFO "  Disabling PCI memory "
1051				       "write and invalidate\n");
1052			}
1053		} else if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1054			printk(KERN_INFO "  PCI memory write & invalidate "
1055			       "enabled by BIOS, enabling counter measures\n");
1056
1057			switch(SMP_CACHE_BYTES) {
1058			case 16:
1059				tmp |= DMA_WRITE_MAX_16;
1060				break;
1061			case 32:
1062				tmp |= DMA_WRITE_MAX_32;
1063				break;
1064			case 64:
1065				tmp |= DMA_WRITE_MAX_64;
1066				break;
1067			case 128:
1068				tmp |= DMA_WRITE_MAX_128;
1069				break;
1070			default:
1071				printk(KERN_INFO "  Cache line size %i not "
1072				       "supported, PCI write and invalidate "
1073				       "disabled\n", SMP_CACHE_BYTES);
1074				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1075				pci_write_config_word(pdev, PCI_COMMAND,
1076						      ap->pci_command);
1077			}
1078		}
1079	}
1080
1081#ifdef __sparc__
1082	/*
1083	 * On this platform, we know what the best dma settings
1084	 * are.  We use 64-byte maximum bursts, because if we
1085	 * burst larger than the cache line size (or even cross
1086	 * a 64byte boundary in a single burst) the UltraSparc
1087	 * PCI controller will disconnect at 64-byte multiples.
1088	 *
1089	 * Read-multiple will be properly enabled above, and when
1090	 * set will give the PCI controller proper hints about
1091	 * prefetching.
1092	 */
1093	tmp &= ~DMA_READ_WRITE_MASK;
1094	tmp |= DMA_READ_MAX_64;
1095	tmp |= DMA_WRITE_MAX_64;
1096#endif
1097#ifdef __alpha__
1098	tmp &= ~DMA_READ_WRITE_MASK;
1099	tmp |= DMA_READ_MAX_128;
1100	/*
1101	 * All the docs say MUST NOT. Well, I did.
1102	 * Nothing terrible happens, if we load wrong size.
1103	 * Bit w&i still works better!
1104	 */
1105	tmp |= DMA_WRITE_MAX_128;
1106#endif
1107	writel(tmp, &regs->PciState);
1108
1109#if 0
1110	/*
1111	 * The Host PCI bus controller driver has to set FBB.
1112	 * If all devices on that PCI bus support FBB, then the controller
1113	 * can enable FBB support in the Host PCI Bus controller (or on
1114	 * the PCI-PCI bridge if that applies).
1115	 * -ggg
1116	 */
1117	/*
1118	 * I have received reports from people having problems when this
1119	 * bit is enabled.
1120	 */
1121	if (!(ap->pci_command & PCI_COMMAND_FAST_BACK)) {
1122		printk(KERN_INFO "  Enabling PCI Fast Back to Back\n");
1123		ap->pci_command |= PCI_COMMAND_FAST_BACK;
1124		pci_write_config_word(pdev, PCI_COMMAND, ap->pci_command);
1125	}
1126#endif
1127
1128	/*
1129	 * Configure DMA attributes.
1130	 */
1131	if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(64))) {
1132		ap->pci_using_dac = 1;
1133	} else if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(32))) {
1134		ap->pci_using_dac = 0;
1135	} else {
1136		ecode = -ENODEV;
1137		goto init_error;
1138	}
1139
1140	/*
1141	 * Initialize the generic info block and the command+event rings
1142	 * and the control blocks for the transmit and receive rings
1143	 * as they need to be setup once and for all.
1144	 */
1145	if (!(info = pci_alloc_consistent(ap->pdev, sizeof(struct ace_info),
1146					  &ap->info_dma))) {
1147		ecode = -EAGAIN;
1148		goto init_error;
1149	}
1150	ap->info = info;
1151
1152	/*
1153	 * Get the memory for the skb rings.
1154	 */
1155	if (!(ap->skb = kmalloc(sizeof(struct ace_skb), GFP_KERNEL))) {
1156		ecode = -EAGAIN;
1157		goto init_error;
1158	}
1159
1160	ecode = request_irq(pdev->irq, ace_interrupt, IRQF_SHARED,
1161			    DRV_NAME, dev);
1162	if (ecode) {
1163		printk(KERN_WARNING "%s: Requested IRQ %d is busy\n",
1164		       DRV_NAME, pdev->irq);
1165		goto init_error;
1166	} else
1167		dev->irq = pdev->irq;
1168
1169#ifdef INDEX_DEBUG
1170	spin_lock_init(&ap->debug_lock);
1171	ap->last_tx = ACE_TX_RING_ENTRIES(ap) - 1;
1172	ap->last_std_rx = 0;
1173	ap->last_mini_rx = 0;
1174#endif
1175
1176	memset(ap->info, 0, sizeof(struct ace_info));
1177	memset(ap->skb, 0, sizeof(struct ace_skb));
1178
1179	ecode = ace_load_firmware(dev);
1180	if (ecode)
1181		goto init_error;
1182
1183	ap->fw_running = 0;
1184
1185	tmp_ptr = ap->info_dma;
1186	writel(tmp_ptr >> 32, &regs->InfoPtrHi);
1187	writel(tmp_ptr & 0xffffffff, &regs->InfoPtrLo);
1188
1189	memset(ap->evt_ring, 0, EVT_RING_ENTRIES * sizeof(struct event));
1190
1191	set_aceaddr(&info->evt_ctrl.rngptr, ap->evt_ring_dma);
1192	info->evt_ctrl.flags = 0;
1193
1194	*(ap->evt_prd) = 0;
1195	wmb();
1196	set_aceaddr(&info->evt_prd_ptr, ap->evt_prd_dma);
1197	writel(0, &regs->EvtCsm);
1198
1199	set_aceaddr(&info->cmd_ctrl.rngptr, 0x100);
1200	info->cmd_ctrl.flags = 0;
1201	info->cmd_ctrl.max_len = 0;
1202
1203	for (i = 0; i < CMD_RING_ENTRIES; i++)
1204		writel(0, &regs->CmdRng[i]);
1205
1206	writel(0, &regs->CmdPrd);
1207	writel(0, &regs->CmdCsm);
1208
1209	tmp_ptr = ap->info_dma;
1210	tmp_ptr += (unsigned long) &(((struct ace_info *)0)->s.stats);
1211	set_aceaddr(&info->stats2_ptr, (dma_addr_t) tmp_ptr);
1212
1213	set_aceaddr(&info->rx_std_ctrl.rngptr, ap->rx_ring_base_dma);
1214	info->rx_std_ctrl.max_len = ACE_STD_BUFSIZE;
1215	info->rx_std_ctrl.flags =
1216	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1217
1218	memset(ap->rx_std_ring, 0,
1219	       RX_STD_RING_ENTRIES * sizeof(struct rx_desc));
1220
1221	for (i = 0; i < RX_STD_RING_ENTRIES; i++)
1222		ap->rx_std_ring[i].flags = BD_FLG_TCP_UDP_SUM;
1223
1224	ap->rx_std_skbprd = 0;
1225	atomic_set(&ap->cur_rx_bufs, 0);
1226
1227	set_aceaddr(&info->rx_jumbo_ctrl.rngptr,
1228		    (ap->rx_ring_base_dma +
1229		     (sizeof(struct rx_desc) * RX_STD_RING_ENTRIES)));
1230	info->rx_jumbo_ctrl.max_len = 0;
1231	info->rx_jumbo_ctrl.flags =
1232	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1233
1234	memset(ap->rx_jumbo_ring, 0,
1235	       RX_JUMBO_RING_ENTRIES * sizeof(struct rx_desc));
1236
1237	for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++)
1238		ap->rx_jumbo_ring[i].flags = BD_FLG_TCP_UDP_SUM | BD_FLG_JUMBO;
1239
1240	ap->rx_jumbo_skbprd = 0;
1241	atomic_set(&ap->cur_jumbo_bufs, 0);
1242
1243	memset(ap->rx_mini_ring, 0,
1244	       RX_MINI_RING_ENTRIES * sizeof(struct rx_desc));
1245
1246	if (ap->version >= 2) {
1247		set_aceaddr(&info->rx_mini_ctrl.rngptr,
1248			    (ap->rx_ring_base_dma +
1249			     (sizeof(struct rx_desc) *
1250			      (RX_STD_RING_ENTRIES +
1251			       RX_JUMBO_RING_ENTRIES))));
1252		info->rx_mini_ctrl.max_len = ACE_MINI_SIZE;
1253		info->rx_mini_ctrl.flags =
1254		  RCB_FLG_TCP_UDP_SUM|RCB_FLG_NO_PSEUDO_HDR|RCB_FLG_VLAN_ASSIST;
1255
1256		for (i = 0; i < RX_MINI_RING_ENTRIES; i++)
1257			ap->rx_mini_ring[i].flags =
1258				BD_FLG_TCP_UDP_SUM | BD_FLG_MINI;
1259	} else {
1260		set_aceaddr(&info->rx_mini_ctrl.rngptr, 0);
1261		info->rx_mini_ctrl.flags = RCB_FLG_RNG_DISABLE;
1262		info->rx_mini_ctrl.max_len = 0;
1263	}
1264
1265	ap->rx_mini_skbprd = 0;
1266	atomic_set(&ap->cur_mini_bufs, 0);
1267
1268	set_aceaddr(&info->rx_return_ctrl.rngptr,
1269		    (ap->rx_ring_base_dma +
1270		     (sizeof(struct rx_desc) *
1271		      (RX_STD_RING_ENTRIES +
1272		       RX_JUMBO_RING_ENTRIES +
1273		       RX_MINI_RING_ENTRIES))));
1274	info->rx_return_ctrl.flags = 0;
1275	info->rx_return_ctrl.max_len = RX_RETURN_RING_ENTRIES;
1276
1277	memset(ap->rx_return_ring, 0,
1278	       RX_RETURN_RING_ENTRIES * sizeof(struct rx_desc));
1279
1280	set_aceaddr(&info->rx_ret_prd_ptr, ap->rx_ret_prd_dma);
1281	*(ap->rx_ret_prd) = 0;
1282
1283	writel(TX_RING_BASE, &regs->WinBase);
1284
1285	if (ACE_IS_TIGON_I(ap)) {
1286		ap->tx_ring = (__force struct tx_desc *) regs->Window;
1287		for (i = 0; i < (TIGON_I_TX_RING_ENTRIES
1288				 * sizeof(struct tx_desc)) / sizeof(u32); i++)
1289			writel(0, (__force void __iomem *)ap->tx_ring  + i * 4);
1290
1291		set_aceaddr(&info->tx_ctrl.rngptr, TX_RING_BASE);
1292	} else {
1293		memset(ap->tx_ring, 0,
1294		       MAX_TX_RING_ENTRIES * sizeof(struct tx_desc));
1295
1296		set_aceaddr(&info->tx_ctrl.rngptr, ap->tx_ring_dma);
1297	}
1298
1299	info->tx_ctrl.max_len = ACE_TX_RING_ENTRIES(ap);
1300	tmp = RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1301
1302	/*
1303	 * The Tigon I does not like having the TX ring in host memory ;-(
1304	 */
1305	if (!ACE_IS_TIGON_I(ap))
1306		tmp |= RCB_FLG_TX_HOST_RING;
1307#if TX_COAL_INTS_ONLY
1308	tmp |= RCB_FLG_COAL_INT_ONLY;
1309#endif
1310	info->tx_ctrl.flags = tmp;
1311
1312	set_aceaddr(&info->tx_csm_ptr, ap->tx_csm_dma);
1313
1314	/*
1315	 * Potential item for tuning parameter
1316	 */
1317#if 0 /* NO */
1318	writel(DMA_THRESH_16W, &regs->DmaReadCfg);
1319	writel(DMA_THRESH_16W, &regs->DmaWriteCfg);
1320#else
1321	writel(DMA_THRESH_8W, &regs->DmaReadCfg);
1322	writel(DMA_THRESH_8W, &regs->DmaWriteCfg);
1323#endif
1324
1325	writel(0, &regs->MaskInt);
1326	writel(1, &regs->IfIdx);
1327#if 0
1328	/*
1329	 * McKinley boxes do not like us fiddling with AssistState
1330	 * this early
1331	 */
1332	writel(1, &regs->AssistState);
1333#endif
1334
1335	writel(DEF_STAT, &regs->TuneStatTicks);
1336	writel(DEF_TRACE, &regs->TuneTrace);
1337
1338	ace_set_rxtx_parms(dev, 0);
1339
1340	if (board_idx == BOARD_IDX_OVERFLOW) {
1341		printk(KERN_WARNING "%s: more than %i NICs detected, "
1342		       "ignoring module parameters!\n",
1343		       ap->name, ACE_MAX_MOD_PARMS);
1344	} else if (board_idx >= 0) {
1345		if (tx_coal_tick[board_idx])
1346			writel(tx_coal_tick[board_idx],
1347			       &regs->TuneTxCoalTicks);
1348		if (max_tx_desc[board_idx])
1349			writel(max_tx_desc[board_idx], &regs->TuneMaxTxDesc);
1350
1351		if (rx_coal_tick[board_idx])
1352			writel(rx_coal_tick[board_idx],
1353			       &regs->TuneRxCoalTicks);
1354		if (max_rx_desc[board_idx])
1355			writel(max_rx_desc[board_idx], &regs->TuneMaxRxDesc);
1356
1357		if (trace[board_idx])
1358			writel(trace[board_idx], &regs->TuneTrace);
1359
1360		if ((tx_ratio[board_idx] > 0) && (tx_ratio[board_idx] < 64))
1361			writel(tx_ratio[board_idx], &regs->TxBufRat);
1362	}
1363
1364	/*
1365	 * Default link parameters
1366	 */
1367	tmp = LNK_ENABLE | LNK_FULL_DUPLEX | LNK_1000MB | LNK_100MB |
1368		LNK_10MB | LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL | LNK_NEGOTIATE;
1369	if(ap->version >= 2)
1370		tmp |= LNK_TX_FLOW_CTL_Y;
1371
1372	/*
1373	 * Override link default parameters
1374	 */
1375	if ((board_idx >= 0) && link_state[board_idx]) {
1376		int option = link_state[board_idx];
1377
1378		tmp = LNK_ENABLE;
1379
1380		if (option & 0x01) {
1381			printk(KERN_INFO "%s: Setting half duplex link\n",
1382			       ap->name);
1383			tmp &= ~LNK_FULL_DUPLEX;
1384		}
1385		if (option & 0x02)
1386			tmp &= ~LNK_NEGOTIATE;
1387		if (option & 0x10)
1388			tmp |= LNK_10MB;
1389		if (option & 0x20)
1390			tmp |= LNK_100MB;
1391		if (option & 0x40)
1392			tmp |= LNK_1000MB;
1393		if ((option & 0x70) == 0) {
1394			printk(KERN_WARNING "%s: No media speed specified, "
1395			       "forcing auto negotiation\n", ap->name);
1396			tmp |= LNK_NEGOTIATE | LNK_1000MB |
1397				LNK_100MB | LNK_10MB;
1398		}
1399		if ((option & 0x100) == 0)
1400			tmp |= LNK_NEG_FCTL;
1401		else
1402			printk(KERN_INFO "%s: Disabling flow control "
1403			       "negotiation\n", ap->name);
1404		if (option & 0x200)
1405			tmp |= LNK_RX_FLOW_CTL_Y;
1406		if ((option & 0x400) && (ap->version >= 2)) {
1407			printk(KERN_INFO "%s: Enabling TX flow control\n",
1408			       ap->name);
1409			tmp |= LNK_TX_FLOW_CTL_Y;
1410		}
1411	}
1412
1413	ap->link = tmp;
1414	writel(tmp, &regs->TuneLink);
1415	if (ap->version >= 2)
1416		writel(tmp, &regs->TuneFastLink);
1417
1418	writel(ap->firmware_start, &regs->Pc);
1419
1420	writel(0, &regs->Mb0Lo);
1421
1422	/*
1423	 * Set tx_csm before we start receiving interrupts, otherwise
1424	 * the interrupt handler might think it is supposed to process
1425	 * tx ints before we are up and running, which may cause a null
1426	 * pointer access in the int handler.
1427	 */
1428	ap->cur_rx = 0;
1429	ap->tx_prd = *(ap->tx_csm) = ap->tx_ret_csm = 0;
1430
1431	wmb();
1432	ace_set_txprd(regs, ap, 0);
1433	writel(0, &regs->RxRetCsm);
1434
1435       /*
1436	* Enable DMA engine now.
1437	* If we do this sooner, Mckinley box pukes.
1438	* I assume it's because Tigon II DMA engine wants to check
1439	* *something* even before the CPU is started.
1440	*/
1441       writel(1, &regs->AssistState);  /* enable DMA */
1442
1443	/*
1444	 * Start the NIC CPU
1445	 */
1446	writel(readl(&regs->CpuCtrl) & ~(CPU_HALT|CPU_TRACE), &regs->CpuCtrl);
1447	readl(&regs->CpuCtrl);
1448
1449	/*
1450	 * Wait for the firmware to spin up - max 3 seconds.
1451	 */
1452	myjif = jiffies + 3 * HZ;
1453	while (time_before(jiffies, myjif) && !ap->fw_running)
1454		cpu_relax();
1455
1456	if (!ap->fw_running) {
1457		printk(KERN_ERR "%s: Firmware NOT running!\n", ap->name);
1458
1459		ace_dump_trace(ap);
1460		writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
1461		readl(&regs->CpuCtrl);
1462
1463		/* aman@sgi.com - account for badly behaving firmware/NIC:
1464		 * - have observed that the NIC may continue to generate
1465		 *   interrupts for some reason; attempt to stop it - halt
1466		 *   second CPU for Tigon II cards, and also clear Mb0
1467		 * - if we're a module, we'll fail to load if this was
1468		 *   the only GbE card in the system => if the kernel does
1469		 *   see an interrupt from the NIC, code to handle it is
1470		 *   gone and OOps! - so free_irq also
1471		 */
1472		if (ap->version >= 2)
1473			writel(readl(&regs->CpuBCtrl) | CPU_HALT,
1474			       &regs->CpuBCtrl);
1475		writel(0, &regs->Mb0Lo);
1476		readl(&regs->Mb0Lo);
1477
1478		ecode = -EBUSY;
1479		goto init_error;
1480	}
1481
1482	/*
1483	 * We load the ring here as there seem to be no way to tell the
1484	 * firmware to wipe the ring without re-initializing it.
1485	 */
1486	if (!test_and_set_bit(0, &ap->std_refill_busy))
1487		ace_load_std_rx_ring(dev, RX_RING_SIZE);
1488	else
1489		printk(KERN_ERR "%s: Someone is busy refilling the RX ring\n",
1490		       ap->name);
1491	if (ap->version >= 2) {
1492		if (!test_and_set_bit(0, &ap->mini_refill_busy))
1493			ace_load_mini_rx_ring(dev, RX_MINI_SIZE);
1494		else
1495			printk(KERN_ERR "%s: Someone is busy refilling "
1496			       "the RX mini ring\n", ap->name);
1497	}
1498	return 0;
1499
1500 init_error:
1501	ace_init_cleanup(dev);
1502	return ecode;
1503}
1504
1505
1506static void ace_set_rxtx_parms(struct net_device *dev, int jumbo)
1507{
1508	struct ace_private *ap = netdev_priv(dev);
1509	struct ace_regs __iomem *regs = ap->regs;
1510	int board_idx = ap->board_idx;
1511
1512	if (board_idx >= 0) {
1513		if (!jumbo) {
1514			if (!tx_coal_tick[board_idx])
1515				writel(DEF_TX_COAL, &regs->TuneTxCoalTicks);
1516			if (!max_tx_desc[board_idx])
1517				writel(DEF_TX_MAX_DESC, &regs->TuneMaxTxDesc);
1518			if (!rx_coal_tick[board_idx])
1519				writel(DEF_RX_COAL, &regs->TuneRxCoalTicks);
1520			if (!max_rx_desc[board_idx])
1521				writel(DEF_RX_MAX_DESC, &regs->TuneMaxRxDesc);
1522			if (!tx_ratio[board_idx])
1523				writel(DEF_TX_RATIO, &regs->TxBufRat);
1524		} else {
1525			if (!tx_coal_tick[board_idx])
1526				writel(DEF_JUMBO_TX_COAL,
1527				       &regs->TuneTxCoalTicks);
1528			if (!max_tx_desc[board_idx])
1529				writel(DEF_JUMBO_TX_MAX_DESC,
1530				       &regs->TuneMaxTxDesc);
1531			if (!rx_coal_tick[board_idx])
1532				writel(DEF_JUMBO_RX_COAL,
1533				       &regs->TuneRxCoalTicks);
1534			if (!max_rx_desc[board_idx])
1535				writel(DEF_JUMBO_RX_MAX_DESC,
1536				       &regs->TuneMaxRxDesc);
1537			if (!tx_ratio[board_idx])
1538				writel(DEF_JUMBO_TX_RATIO, &regs->TxBufRat);
1539		}
1540	}
1541}
1542
1543
1544static void ace_watchdog(struct net_device *data)
1545{
1546	struct net_device *dev = data;
1547	struct ace_private *ap = netdev_priv(dev);
1548	struct ace_regs __iomem *regs = ap->regs;
1549
1550	/*
1551	 * We haven't received a stats update event for more than 2.5
1552	 * seconds and there is data in the transmit queue, thus we
1553	 * assume the card is stuck.
1554	 */
1555	if (*ap->tx_csm != ap->tx_ret_csm) {
1556		printk(KERN_WARNING "%s: Transmitter is stuck, %08x\n",
1557		       dev->name, (unsigned int)readl(&regs->HostCtrl));
1558		/* This can happen due to ieee flow control. */
1559	} else {
1560		printk(KERN_DEBUG "%s: BUG... transmitter died. Kicking it.\n",
1561		       dev->name);
1562#if 0
1563		netif_wake_queue(dev);
1564#endif
1565	}
1566}
1567
1568
1569static void ace_tasklet(unsigned long arg)
1570{
1571	struct net_device *dev = (struct net_device *) arg;
1572	struct ace_private *ap = netdev_priv(dev);
1573	int cur_size;
1574
1575	cur_size = atomic_read(&ap->cur_rx_bufs);
1576	if ((cur_size < RX_LOW_STD_THRES) &&
1577	    !test_and_set_bit(0, &ap->std_refill_busy)) {
1578#ifdef DEBUG
1579		printk("refilling buffers (current %i)\n", cur_size);
1580#endif
1581		ace_load_std_rx_ring(dev, RX_RING_SIZE - cur_size);
1582	}
1583
1584	if (ap->version >= 2) {
1585		cur_size = atomic_read(&ap->cur_mini_bufs);
1586		if ((cur_size < RX_LOW_MINI_THRES) &&
1587		    !test_and_set_bit(0, &ap->mini_refill_busy)) {
1588#ifdef DEBUG
1589			printk("refilling mini buffers (current %i)\n",
1590			       cur_size);
1591#endif
1592			ace_load_mini_rx_ring(dev, RX_MINI_SIZE - cur_size);
1593		}
1594	}
1595
1596	cur_size = atomic_read(&ap->cur_jumbo_bufs);
1597	if (ap->jumbo && (cur_size < RX_LOW_JUMBO_THRES) &&
1598	    !test_and_set_bit(0, &ap->jumbo_refill_busy)) {
1599#ifdef DEBUG
1600		printk("refilling jumbo buffers (current %i)\n", cur_size);
1601#endif
1602		ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE - cur_size);
1603	}
1604	ap->tasklet_pending = 0;
1605}
1606
1607
1608/*
1609 * Copy the contents of the NIC's trace buffer to kernel memory.
1610 */
1611static void ace_dump_trace(struct ace_private *ap)
1612{
1613#if 0
1614	if (!ap->trace_buf)
1615		if (!(ap->trace_buf = kmalloc(ACE_TRACE_SIZE, GFP_KERNEL)))
1616		    return;
1617#endif
1618}
1619
1620
1621/*
1622 * Load the standard rx ring.
1623 *
1624 * Loading rings is safe without holding the spin lock since this is
1625 * done only before the device is enabled, thus no interrupts are
1626 * generated and by the interrupt handler/tasklet handler.
1627 */
1628static void ace_load_std_rx_ring(struct net_device *dev, int nr_bufs)
1629{
1630	struct ace_private *ap = netdev_priv(dev);
1631	struct ace_regs __iomem *regs = ap->regs;
1632	short i, idx;
1633
1634
1635	prefetchw(&ap->cur_rx_bufs);
1636
1637	idx = ap->rx_std_skbprd;
1638
1639	for (i = 0; i < nr_bufs; i++) {
1640		struct sk_buff *skb;
1641		struct rx_desc *rd;
1642		dma_addr_t mapping;
1643
1644		skb = netdev_alloc_skb_ip_align(dev, ACE_STD_BUFSIZE);
1645		if (!skb)
1646			break;
1647
1648		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1649				       offset_in_page(skb->data),
1650				       ACE_STD_BUFSIZE,
1651				       PCI_DMA_FROMDEVICE);
1652		ap->skb->rx_std_skbuff[idx].skb = skb;
1653		dma_unmap_addr_set(&ap->skb->rx_std_skbuff[idx],
1654				   mapping, mapping);
1655
1656		rd = &ap->rx_std_ring[idx];
1657		set_aceaddr(&rd->addr, mapping);
1658		rd->size = ACE_STD_BUFSIZE;
1659		rd->idx = idx;
1660		idx = (idx + 1) % RX_STD_RING_ENTRIES;
1661	}
1662
1663	if (!i)
1664		goto error_out;
1665
1666	atomic_add(i, &ap->cur_rx_bufs);
1667	ap->rx_std_skbprd = idx;
1668
1669	if (ACE_IS_TIGON_I(ap)) {
1670		struct cmd cmd;
1671		cmd.evt = C_SET_RX_PRD_IDX;
1672		cmd.code = 0;
1673		cmd.idx = ap->rx_std_skbprd;
1674		ace_issue_cmd(regs, &cmd);
1675	} else {
1676		writel(idx, &regs->RxStdPrd);
1677		wmb();
1678	}
1679
1680 out:
1681	clear_bit(0, &ap->std_refill_busy);
1682	return;
1683
1684 error_out:
1685	printk(KERN_INFO "Out of memory when allocating "
1686	       "standard receive buffers\n");
1687	goto out;
1688}
1689
1690
1691static void ace_load_mini_rx_ring(struct net_device *dev, int nr_bufs)
1692{
1693	struct ace_private *ap = netdev_priv(dev);
1694	struct ace_regs __iomem *regs = ap->regs;
1695	short i, idx;
1696
1697	prefetchw(&ap->cur_mini_bufs);
1698
1699	idx = ap->rx_mini_skbprd;
1700	for (i = 0; i < nr_bufs; i++) {
1701		struct sk_buff *skb;
1702		struct rx_desc *rd;
1703		dma_addr_t mapping;
1704
1705		skb = netdev_alloc_skb_ip_align(dev, ACE_MINI_BUFSIZE);
1706		if (!skb)
1707			break;
1708
1709		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1710				       offset_in_page(skb->data),
1711				       ACE_MINI_BUFSIZE,
1712				       PCI_DMA_FROMDEVICE);
1713		ap->skb->rx_mini_skbuff[idx].skb = skb;
1714		dma_unmap_addr_set(&ap->skb->rx_mini_skbuff[idx],
1715				   mapping, mapping);
1716
1717		rd = &ap->rx_mini_ring[idx];
1718		set_aceaddr(&rd->addr, mapping);
1719		rd->size = ACE_MINI_BUFSIZE;
1720		rd->idx = idx;
1721		idx = (idx + 1) % RX_MINI_RING_ENTRIES;
1722	}
1723
1724	if (!i)
1725		goto error_out;
1726
1727	atomic_add(i, &ap->cur_mini_bufs);
1728
1729	ap->rx_mini_skbprd = idx;
1730
1731	writel(idx, &regs->RxMiniPrd);
1732	wmb();
1733
1734 out:
1735	clear_bit(0, &ap->mini_refill_busy);
1736	return;
1737 error_out:
1738	printk(KERN_INFO "Out of memory when allocating "
1739	       "mini receive buffers\n");
1740	goto out;
1741}
1742
1743
1744/*
1745 * Load the jumbo rx ring, this may happen at any time if the MTU
1746 * is changed to a value > 1500.
1747 */
1748static void ace_load_jumbo_rx_ring(struct net_device *dev, int nr_bufs)
1749{
1750	struct ace_private *ap = netdev_priv(dev);
1751	struct ace_regs __iomem *regs = ap->regs;
1752	short i, idx;
1753
1754	idx = ap->rx_jumbo_skbprd;
1755
1756	for (i = 0; i < nr_bufs; i++) {
1757		struct sk_buff *skb;
1758		struct rx_desc *rd;
1759		dma_addr_t mapping;
1760
1761		skb = netdev_alloc_skb_ip_align(dev, ACE_JUMBO_BUFSIZE);
1762		if (!skb)
1763			break;
1764
1765		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1766				       offset_in_page(skb->data),
1767				       ACE_JUMBO_BUFSIZE,
1768				       PCI_DMA_FROMDEVICE);
1769		ap->skb->rx_jumbo_skbuff[idx].skb = skb;
1770		dma_unmap_addr_set(&ap->skb->rx_jumbo_skbuff[idx],
1771				   mapping, mapping);
1772
1773		rd = &ap->rx_jumbo_ring[idx];
1774		set_aceaddr(&rd->addr, mapping);
1775		rd->size = ACE_JUMBO_BUFSIZE;
1776		rd->idx = idx;
1777		idx = (idx + 1) % RX_JUMBO_RING_ENTRIES;
1778	}
1779
1780	if (!i)
1781		goto error_out;
1782
1783	atomic_add(i, &ap->cur_jumbo_bufs);
1784	ap->rx_jumbo_skbprd = idx;
1785
1786	if (ACE_IS_TIGON_I(ap)) {
1787		struct cmd cmd;
1788		cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1789		cmd.code = 0;
1790		cmd.idx = ap->rx_jumbo_skbprd;
1791		ace_issue_cmd(regs, &cmd);
1792	} else {
1793		writel(idx, &regs->RxJumboPrd);
1794		wmb();
1795	}
1796
1797 out:
1798	clear_bit(0, &ap->jumbo_refill_busy);
1799	return;
1800 error_out:
1801	if (net_ratelimit())
1802		printk(KERN_INFO "Out of memory when allocating "
1803		       "jumbo receive buffers\n");
1804	goto out;
1805}
1806
1807
1808/*
1809 * All events are considered to be slow (RX/TX ints do not generate
1810 * events) and are handled here, outside the main interrupt handler,
1811 * to reduce the size of the handler.
1812 */
1813static u32 ace_handle_event(struct net_device *dev, u32 evtcsm, u32 evtprd)
1814{
1815	struct ace_private *ap;
1816
1817	ap = netdev_priv(dev);
1818
1819	while (evtcsm != evtprd) {
1820		switch (ap->evt_ring[evtcsm].evt) {
1821		case E_FW_RUNNING:
1822			printk(KERN_INFO "%s: Firmware up and running\n",
1823			       ap->name);
1824			ap->fw_running = 1;
1825			wmb();
1826			break;
1827		case E_STATS_UPDATED:
1828			break;
1829		case E_LNK_STATE:
1830		{
1831			u16 code = ap->evt_ring[evtcsm].code;
1832			switch (code) {
1833			case E_C_LINK_UP:
1834			{
1835				u32 state = readl(&ap->regs->GigLnkState);
1836				printk(KERN_WARNING "%s: Optical link UP "
1837				       "(%s Duplex, Flow Control: %s%s)\n",
1838				       ap->name,
1839				       state & LNK_FULL_DUPLEX ? "Full":"Half",
1840				       state & LNK_TX_FLOW_CTL_Y ? "TX " : "",
1841				       state & LNK_RX_FLOW_CTL_Y ? "RX" : "");
1842				break;
1843			}
1844			case E_C_LINK_DOWN:
1845				printk(KERN_WARNING "%s: Optical link DOWN\n",
1846				       ap->name);
1847				break;
1848			case E_C_LINK_10_100:
1849				printk(KERN_WARNING "%s: 10/100BaseT link "
1850				       "UP\n", ap->name);
1851				break;
1852			default:
1853				printk(KERN_ERR "%s: Unknown optical link "
1854				       "state %02x\n", ap->name, code);
1855			}
1856			break;
1857		}
1858		case E_ERROR:
1859			switch(ap->evt_ring[evtcsm].code) {
1860			case E_C_ERR_INVAL_CMD:
1861				printk(KERN_ERR "%s: invalid command error\n",
1862				       ap->name);
1863				break;
1864			case E_C_ERR_UNIMP_CMD:
1865				printk(KERN_ERR "%s: unimplemented command "
1866				       "error\n", ap->name);
1867				break;
1868			case E_C_ERR_BAD_CFG:
1869				printk(KERN_ERR "%s: bad config error\n",
1870				       ap->name);
1871				break;
1872			default:
1873				printk(KERN_ERR "%s: unknown error %02x\n",
1874				       ap->name, ap->evt_ring[evtcsm].code);
1875			}
1876			break;
1877		case E_RESET_JUMBO_RNG:
1878		{
1879			int i;
1880			for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
1881				if (ap->skb->rx_jumbo_skbuff[i].skb) {
1882					ap->rx_jumbo_ring[i].size = 0;
1883					set_aceaddr(&ap->rx_jumbo_ring[i].addr, 0);
1884					dev_kfree_skb(ap->skb->rx_jumbo_skbuff[i].skb);
1885					ap->skb->rx_jumbo_skbuff[i].skb = NULL;
1886				}
1887			}
1888
1889 			if (ACE_IS_TIGON_I(ap)) {
1890 				struct cmd cmd;
1891 				cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1892 				cmd.code = 0;
1893 				cmd.idx = 0;
1894 				ace_issue_cmd(ap->regs, &cmd);
1895 			} else {
1896 				writel(0, &((ap->regs)->RxJumboPrd));
1897 				wmb();
1898 			}
1899
1900			ap->jumbo = 0;
1901			ap->rx_jumbo_skbprd = 0;
1902			printk(KERN_INFO "%s: Jumbo ring flushed\n",
1903			       ap->name);
1904			clear_bit(0, &ap->jumbo_refill_busy);
1905			break;
1906		}
1907		default:
1908			printk(KERN_ERR "%s: Unhandled event 0x%02x\n",
1909			       ap->name, ap->evt_ring[evtcsm].evt);
1910		}
1911		evtcsm = (evtcsm + 1) % EVT_RING_ENTRIES;
1912	}
1913
1914	return evtcsm;
1915}
1916
1917
1918static void ace_rx_int(struct net_device *dev, u32 rxretprd, u32 rxretcsm)
1919{
1920	struct ace_private *ap = netdev_priv(dev);
1921	u32 idx;
1922	int mini_count = 0, std_count = 0;
1923
1924	idx = rxretcsm;
1925
1926	prefetchw(&ap->cur_rx_bufs);
1927	prefetchw(&ap->cur_mini_bufs);
1928
1929	while (idx != rxretprd) {
1930		struct ring_info *rip;
1931		struct sk_buff *skb;
1932		struct rx_desc *rxdesc, *retdesc;
1933		u32 skbidx;
1934		int bd_flags, desc_type, mapsize;
1935		u16 csum;
1936
1937
1938		/* make sure the rx descriptor isn't read before rxretprd */
1939		if (idx == rxretcsm)
1940			rmb();
1941
1942		retdesc = &ap->rx_return_ring[idx];
1943		skbidx = retdesc->idx;
1944		bd_flags = retdesc->flags;
1945		desc_type = bd_flags & (BD_FLG_JUMBO | BD_FLG_MINI);
1946
1947		switch(desc_type) {
1948			/*
1949			 * Normal frames do not have any flags set
1950			 *
1951			 * Mini and normal frames arrive frequently,
1952			 * so use a local counter to avoid doing
1953			 * atomic operations for each packet arriving.
1954			 */
1955		case 0:
1956			rip = &ap->skb->rx_std_skbuff[skbidx];
1957			mapsize = ACE_STD_BUFSIZE;
1958			rxdesc = &ap->rx_std_ring[skbidx];
1959			std_count++;
1960			break;
1961		case BD_FLG_JUMBO:
1962			rip = &ap->skb->rx_jumbo_skbuff[skbidx];
1963			mapsize = ACE_JUMBO_BUFSIZE;
1964			rxdesc = &ap->rx_jumbo_ring[skbidx];
1965			atomic_dec(&ap->cur_jumbo_bufs);
1966			break;
1967		case BD_FLG_MINI:
1968			rip = &ap->skb->rx_mini_skbuff[skbidx];
1969			mapsize = ACE_MINI_BUFSIZE;
1970			rxdesc = &ap->rx_mini_ring[skbidx];
1971			mini_count++;
1972			break;
1973		default:
1974			printk(KERN_INFO "%s: unknown frame type (0x%02x) "
1975			       "returned by NIC\n", dev->name,
1976			       retdesc->flags);
1977			goto error;
1978		}
1979
1980		skb = rip->skb;
1981		rip->skb = NULL;
1982		pci_unmap_page(ap->pdev,
1983			       dma_unmap_addr(rip, mapping),
1984			       mapsize,
1985			       PCI_DMA_FROMDEVICE);
1986		skb_put(skb, retdesc->size);
1987
1988		/*
1989		 * Fly baby, fly!
1990		 */
1991		csum = retdesc->tcp_udp_csum;
1992
1993		skb->protocol = eth_type_trans(skb, dev);
1994
1995		/*
1996		 * Instead of forcing the poor tigon mips cpu to calculate
1997		 * pseudo hdr checksum, we do this ourselves.
1998		 */
1999		if (bd_flags & BD_FLG_TCP_UDP_SUM) {
2000			skb->csum = htons(csum);
2001			skb->ip_summed = CHECKSUM_COMPLETE;
2002		} else {
2003			skb_checksum_none_assert(skb);
2004		}
2005
2006		/* send it up */
2007		if ((bd_flags & BD_FLG_VLAN_TAG))
2008			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), retdesc->vlan);
2009		netif_rx(skb);
2010
2011		dev->stats.rx_packets++;
2012		dev->stats.rx_bytes += retdesc->size;
2013
2014		idx = (idx + 1) % RX_RETURN_RING_ENTRIES;
2015	}
2016
2017	atomic_sub(std_count, &ap->cur_rx_bufs);
2018	if (!ACE_IS_TIGON_I(ap))
2019		atomic_sub(mini_count, &ap->cur_mini_bufs);
2020
2021 out:
2022	/*
2023	 * According to the documentation RxRetCsm is obsolete with
2024	 * the 12.3.x Firmware - my Tigon I NICs seem to disagree!
2025	 */
2026	if (ACE_IS_TIGON_I(ap)) {
2027		writel(idx, &ap->regs->RxRetCsm);
2028	}
2029	ap->cur_rx = idx;
2030
2031	return;
2032 error:
2033	idx = rxretprd;
2034	goto out;
2035}
2036
2037
2038static inline void ace_tx_int(struct net_device *dev,
2039			      u32 txcsm, u32 idx)
2040{
2041	struct ace_private *ap = netdev_priv(dev);
2042
2043	do {
2044		struct sk_buff *skb;
2045		struct tx_ring_info *info;
2046
2047		info = ap->skb->tx_skbuff + idx;
2048		skb = info->skb;
2049
2050		if (dma_unmap_len(info, maplen)) {
2051			pci_unmap_page(ap->pdev, dma_unmap_addr(info, mapping),
2052				       dma_unmap_len(info, maplen),
2053				       PCI_DMA_TODEVICE);
2054			dma_unmap_len_set(info, maplen, 0);
2055		}
2056
2057		if (skb) {
2058			dev->stats.tx_packets++;
2059			dev->stats.tx_bytes += skb->len;
2060			dev_kfree_skb_irq(skb);
2061			info->skb = NULL;
2062		}
2063
2064		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2065	} while (idx != txcsm);
2066
2067	if (netif_queue_stopped(dev))
2068		netif_wake_queue(dev);
2069
2070	wmb();
2071	ap->tx_ret_csm = txcsm;
2072
2073	/* So... tx_ret_csm is advanced _after_ check for device wakeup.
2074	 *
2075	 * We could try to make it before. In this case we would get
2076	 * the following race condition: hard_start_xmit on other cpu
2077	 * enters after we advanced tx_ret_csm and fills space,
2078	 * which we have just freed, so that we make illegal device wakeup.
2079	 * There is no good way to workaround this (at entry
2080	 * to ace_start_xmit detects this condition and prevents
2081	 * ring corruption, but it is not a good workaround.)
2082	 *
2083	 * When tx_ret_csm is advanced after, we wake up device _only_
2084	 * if we really have some space in ring (though the core doing
2085	 * hard_start_xmit can see full ring for some period and has to
2086	 * synchronize.) Superb.
2087	 * BUT! We get another subtle race condition. hard_start_xmit
2088	 * may think that ring is full between wakeup and advancing
2089	 * tx_ret_csm and will stop device instantly! It is not so bad.
2090	 * We are guaranteed that there is something in ring, so that
2091	 * the next irq will resume transmission. To speedup this we could
2092	 * mark descriptor, which closes ring with BD_FLG_COAL_NOW
2093	 * (see ace_start_xmit).
2094	 *
2095	 * Well, this dilemma exists in all lock-free devices.
2096	 * We, following scheme used in drivers by Donald Becker,
2097	 * select the least dangerous.
2098	 *							--ANK
2099	 */
2100}
2101
2102
2103static irqreturn_t ace_interrupt(int irq, void *dev_id)
2104{
2105	struct net_device *dev = (struct net_device *)dev_id;
2106	struct ace_private *ap = netdev_priv(dev);
2107	struct ace_regs __iomem *regs = ap->regs;
2108	u32 idx;
2109	u32 txcsm, rxretcsm, rxretprd;
2110	u32 evtcsm, evtprd;
2111
2112	/*
2113	 * In case of PCI shared interrupts or spurious interrupts,
2114	 * we want to make sure it is actually our interrupt before
2115	 * spending any time in here.
2116	 */
2117	if (!(readl(&regs->HostCtrl) & IN_INT))
2118		return IRQ_NONE;
2119
2120	/*
2121	 * ACK intr now. Otherwise we will lose updates to rx_ret_prd,
2122	 * which happened _after_ rxretprd = *ap->rx_ret_prd; but before
2123	 * writel(0, &regs->Mb0Lo).
2124	 *
2125	 * "IRQ avoidance" recommended in docs applies to IRQs served
2126	 * threads and it is wrong even for that case.
2127	 */
2128	writel(0, &regs->Mb0Lo);
2129	readl(&regs->Mb0Lo);
2130
2131	/*
2132	 * There is no conflict between transmit handling in
2133	 * start_xmit and receive processing, thus there is no reason
2134	 * to take a spin lock for RX handling. Wait until we start
2135	 * working on the other stuff - hey we don't need a spin lock
2136	 * anymore.
2137	 */
2138	rxretprd = *ap->rx_ret_prd;
2139	rxretcsm = ap->cur_rx;
2140
2141	if (rxretprd != rxretcsm)
2142		ace_rx_int(dev, rxretprd, rxretcsm);
2143
2144	txcsm = *ap->tx_csm;
2145	idx = ap->tx_ret_csm;
2146
2147	if (txcsm != idx) {
2148		/*
2149		 * If each skb takes only one descriptor this check degenerates
2150		 * to identity, because new space has just been opened.
2151		 * But if skbs are fragmented we must check that this index
2152		 * update releases enough of space, otherwise we just
2153		 * wait for device to make more work.
2154		 */
2155		if (!tx_ring_full(ap, txcsm, ap->tx_prd))
2156			ace_tx_int(dev, txcsm, idx);
2157	}
2158
2159	evtcsm = readl(&regs->EvtCsm);
2160	evtprd = *ap->evt_prd;
2161
2162	if (evtcsm != evtprd) {
2163		evtcsm = ace_handle_event(dev, evtcsm, evtprd);
2164		writel(evtcsm, &regs->EvtCsm);
2165	}
2166
2167	/*
2168	 * This has to go last in the interrupt handler and run with
2169	 * the spin lock released ... what lock?
2170	 */
2171	if (netif_running(dev)) {
2172		int cur_size;
2173		int run_tasklet = 0;
2174
2175		cur_size = atomic_read(&ap->cur_rx_bufs);
2176		if (cur_size < RX_LOW_STD_THRES) {
2177			if ((cur_size < RX_PANIC_STD_THRES) &&
2178			    !test_and_set_bit(0, &ap->std_refill_busy)) {
2179#ifdef DEBUG
2180				printk("low on std buffers %i\n", cur_size);
2181#endif
2182				ace_load_std_rx_ring(dev,
2183						     RX_RING_SIZE - cur_size);
2184			} else
2185				run_tasklet = 1;
2186		}
2187
2188		if (!ACE_IS_TIGON_I(ap)) {
2189			cur_size = atomic_read(&ap->cur_mini_bufs);
2190			if (cur_size < RX_LOW_MINI_THRES) {
2191				if ((cur_size < RX_PANIC_MINI_THRES) &&
2192				    !test_and_set_bit(0,
2193						      &ap->mini_refill_busy)) {
2194#ifdef DEBUG
2195					printk("low on mini buffers %i\n",
2196					       cur_size);
2197#endif
2198					ace_load_mini_rx_ring(dev,
2199							      RX_MINI_SIZE - cur_size);
2200				} else
2201					run_tasklet = 1;
2202			}
2203		}
2204
2205		if (ap->jumbo) {
2206			cur_size = atomic_read(&ap->cur_jumbo_bufs);
2207			if (cur_size < RX_LOW_JUMBO_THRES) {
2208				if ((cur_size < RX_PANIC_JUMBO_THRES) &&
2209				    !test_and_set_bit(0,
2210						      &ap->jumbo_refill_busy)){
2211#ifdef DEBUG
2212					printk("low on jumbo buffers %i\n",
2213					       cur_size);
2214#endif
2215					ace_load_jumbo_rx_ring(dev,
2216							       RX_JUMBO_SIZE - cur_size);
2217				} else
2218					run_tasklet = 1;
2219			}
2220		}
2221		if (run_tasklet && !ap->tasklet_pending) {
2222			ap->tasklet_pending = 1;
2223			tasklet_schedule(&ap->ace_tasklet);
2224		}
2225	}
2226
2227	return IRQ_HANDLED;
2228}
2229
2230static int ace_open(struct net_device *dev)
2231{
2232	struct ace_private *ap = netdev_priv(dev);
2233	struct ace_regs __iomem *regs = ap->regs;
2234	struct cmd cmd;
2235
2236	if (!(ap->fw_running)) {
2237		printk(KERN_WARNING "%s: Firmware not running!\n", dev->name);
2238		return -EBUSY;
2239	}
2240
2241	writel(dev->mtu + ETH_HLEN + 4, &regs->IfMtu);
2242
2243	cmd.evt = C_CLEAR_STATS;
2244	cmd.code = 0;
2245	cmd.idx = 0;
2246	ace_issue_cmd(regs, &cmd);
2247
2248	cmd.evt = C_HOST_STATE;
2249	cmd.code = C_C_STACK_UP;
2250	cmd.idx = 0;
2251	ace_issue_cmd(regs, &cmd);
2252
2253	if (ap->jumbo &&
2254	    !test_and_set_bit(0, &ap->jumbo_refill_busy))
2255		ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE);
2256
2257	if (dev->flags & IFF_PROMISC) {
2258		cmd.evt = C_SET_PROMISC_MODE;
2259		cmd.code = C_C_PROMISC_ENABLE;
2260		cmd.idx = 0;
2261		ace_issue_cmd(regs, &cmd);
2262
2263		ap->promisc = 1;
2264	}else
2265		ap->promisc = 0;
2266	ap->mcast_all = 0;
2267
2268#if 0
2269	cmd.evt = C_LNK_NEGOTIATION;
2270	cmd.code = 0;
2271	cmd.idx = 0;
2272	ace_issue_cmd(regs, &cmd);
2273#endif
2274
2275	netif_start_queue(dev);
2276
2277	/*
2278	 * Setup the bottom half rx ring refill handler
2279	 */
2280	tasklet_init(&ap->ace_tasklet, ace_tasklet, (unsigned long)dev);
2281	return 0;
2282}
2283
2284
2285static int ace_close(struct net_device *dev)
2286{
2287	struct ace_private *ap = netdev_priv(dev);
2288	struct ace_regs __iomem *regs = ap->regs;
2289	struct cmd cmd;
2290	unsigned long flags;
2291	short i;
2292
2293	/*
2294	 * Without (or before) releasing irq and stopping hardware, this
2295	 * is an absolute non-sense, by the way. It will be reset instantly
2296	 * by the first irq.
2297	 */
2298	netif_stop_queue(dev);
2299
2300
2301	if (ap->promisc) {
2302		cmd.evt = C_SET_PROMISC_MODE;
2303		cmd.code = C_C_PROMISC_DISABLE;
2304		cmd.idx = 0;
2305		ace_issue_cmd(regs, &cmd);
2306		ap->promisc = 0;
2307	}
2308
2309	cmd.evt = C_HOST_STATE;
2310	cmd.code = C_C_STACK_DOWN;
2311	cmd.idx = 0;
2312	ace_issue_cmd(regs, &cmd);
2313
2314	tasklet_kill(&ap->ace_tasklet);
2315
2316	/*
2317	 * Make sure one CPU is not processing packets while
2318	 * buffers are being released by another.
2319	 */
2320
2321	local_irq_save(flags);
2322	ace_mask_irq(dev);
2323
2324	for (i = 0; i < ACE_TX_RING_ENTRIES(ap); i++) {
2325		struct sk_buff *skb;
2326		struct tx_ring_info *info;
2327
2328		info = ap->skb->tx_skbuff + i;
2329		skb = info->skb;
2330
2331		if (dma_unmap_len(info, maplen)) {
2332			if (ACE_IS_TIGON_I(ap)) {
2333				/* NB: TIGON_1 is special, tx_ring is in io space */
2334				struct tx_desc __iomem *tx;
2335				tx = (__force struct tx_desc __iomem *) &ap->tx_ring[i];
2336				writel(0, &tx->addr.addrhi);
2337				writel(0, &tx->addr.addrlo);
2338				writel(0, &tx->flagsize);
2339			} else
2340				memset(ap->tx_ring + i, 0,
2341				       sizeof(struct tx_desc));
2342			pci_unmap_page(ap->pdev, dma_unmap_addr(info, mapping),
2343				       dma_unmap_len(info, maplen),
2344				       PCI_DMA_TODEVICE);
2345			dma_unmap_len_set(info, maplen, 0);
2346		}
2347		if (skb) {
2348			dev_kfree_skb(skb);
2349			info->skb = NULL;
2350		}
2351	}
2352
2353	if (ap->jumbo) {
2354		cmd.evt = C_RESET_JUMBO_RNG;
2355		cmd.code = 0;
2356		cmd.idx = 0;
2357		ace_issue_cmd(regs, &cmd);
2358	}
2359
2360	ace_unmask_irq(dev);
2361	local_irq_restore(flags);
2362
2363	return 0;
2364}
2365
2366
2367static inline dma_addr_t
2368ace_map_tx_skb(struct ace_private *ap, struct sk_buff *skb,
2369	       struct sk_buff *tail, u32 idx)
2370{
2371	dma_addr_t mapping;
2372	struct tx_ring_info *info;
2373
2374	mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
2375			       offset_in_page(skb->data),
2376			       skb->len, PCI_DMA_TODEVICE);
2377
2378	info = ap->skb->tx_skbuff + idx;
2379	info->skb = tail;
2380	dma_unmap_addr_set(info, mapping, mapping);
2381	dma_unmap_len_set(info, maplen, skb->len);
2382	return mapping;
2383}
2384
2385
2386static inline void
2387ace_load_tx_bd(struct ace_private *ap, struct tx_desc *desc, u64 addr,
2388	       u32 flagsize, u32 vlan_tag)
2389{
2390#if !USE_TX_COAL_NOW
2391	flagsize &= ~BD_FLG_COAL_NOW;
2392#endif
2393
2394	if (ACE_IS_TIGON_I(ap)) {
2395		struct tx_desc __iomem *io = (__force struct tx_desc __iomem *) desc;
2396		writel(addr >> 32, &io->addr.addrhi);
2397		writel(addr & 0xffffffff, &io->addr.addrlo);
2398		writel(flagsize, &io->flagsize);
2399		writel(vlan_tag, &io->vlanres);
2400	} else {
2401		desc->addr.addrhi = addr >> 32;
2402		desc->addr.addrlo = addr;
2403		desc->flagsize = flagsize;
2404		desc->vlanres = vlan_tag;
2405	}
2406}
2407
2408
2409static netdev_tx_t ace_start_xmit(struct sk_buff *skb,
2410				  struct net_device *dev)
2411{
2412	struct ace_private *ap = netdev_priv(dev);
2413	struct ace_regs __iomem *regs = ap->regs;
2414	struct tx_desc *desc;
2415	u32 idx, flagsize;
2416	unsigned long maxjiff = jiffies + 3*HZ;
2417
2418restart:
2419	idx = ap->tx_prd;
2420
2421	if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2422		goto overflow;
2423
2424	if (!skb_shinfo(skb)->nr_frags)	{
2425		dma_addr_t mapping;
2426		u32 vlan_tag = 0;
2427
2428		mapping = ace_map_tx_skb(ap, skb, skb, idx);
2429		flagsize = (skb->len << 16) | (BD_FLG_END);
2430		if (skb->ip_summed == CHECKSUM_PARTIAL)
2431			flagsize |= BD_FLG_TCP_UDP_SUM;
2432		if (skb_vlan_tag_present(skb)) {
2433			flagsize |= BD_FLG_VLAN_TAG;
2434			vlan_tag = skb_vlan_tag_get(skb);
2435		}
2436		desc = ap->tx_ring + idx;
2437		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2438
2439		/* Look at ace_tx_int for explanations. */
2440		if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2441			flagsize |= BD_FLG_COAL_NOW;
2442
2443		ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2444	} else {
2445		dma_addr_t mapping;
2446		u32 vlan_tag = 0;
2447		int i, len = 0;
2448
2449		mapping = ace_map_tx_skb(ap, skb, NULL, idx);
2450		flagsize = (skb_headlen(skb) << 16);
2451		if (skb->ip_summed == CHECKSUM_PARTIAL)
2452			flagsize |= BD_FLG_TCP_UDP_SUM;
2453		if (skb_vlan_tag_present(skb)) {
2454			flagsize |= BD_FLG_VLAN_TAG;
2455			vlan_tag = skb_vlan_tag_get(skb);
2456		}
2457
2458		ace_load_tx_bd(ap, ap->tx_ring + idx, mapping, flagsize, vlan_tag);
2459
2460		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2461
2462		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2463			const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2464			struct tx_ring_info *info;
2465
2466			len += skb_frag_size(frag);
2467			info = ap->skb->tx_skbuff + idx;
2468			desc = ap->tx_ring + idx;
2469
2470			mapping = skb_frag_dma_map(&ap->pdev->dev, frag, 0,
2471						   skb_frag_size(frag),
2472						   DMA_TO_DEVICE);
2473
2474			flagsize = skb_frag_size(frag) << 16;
2475			if (skb->ip_summed == CHECKSUM_PARTIAL)
2476				flagsize |= BD_FLG_TCP_UDP_SUM;
2477			idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2478
2479			if (i == skb_shinfo(skb)->nr_frags - 1) {
2480				flagsize |= BD_FLG_END;
2481				if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2482					flagsize |= BD_FLG_COAL_NOW;
2483
2484				/*
2485				 * Only the last fragment frees
2486				 * the skb!
2487				 */
2488				info->skb = skb;
2489			} else {
2490				info->skb = NULL;
2491			}
2492			dma_unmap_addr_set(info, mapping, mapping);
2493			dma_unmap_len_set(info, maplen, skb_frag_size(frag));
2494			ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2495		}
2496	}
2497
2498 	wmb();
2499 	ap->tx_prd = idx;
2500 	ace_set_txprd(regs, ap, idx);
2501
2502	if (flagsize & BD_FLG_COAL_NOW) {
2503		netif_stop_queue(dev);
2504
2505		/*
2506		 * A TX-descriptor producer (an IRQ) might have gotten
2507		 * between, making the ring free again. Since xmit is
2508		 * serialized, this is the only situation we have to
2509		 * re-test.
2510		 */
2511		if (!tx_ring_full(ap, ap->tx_ret_csm, idx))
2512			netif_wake_queue(dev);
2513	}
2514
2515	return NETDEV_TX_OK;
2516
2517overflow:
2518	/*
2519	 * This race condition is unavoidable with lock-free drivers.
2520	 * We wake up the queue _before_ tx_prd is advanced, so that we can
2521	 * enter hard_start_xmit too early, while tx ring still looks closed.
2522	 * This happens ~1-4 times per 100000 packets, so that we can allow
2523	 * to loop syncing to other CPU. Probably, we need an additional
2524	 * wmb() in ace_tx_intr as well.
2525	 *
2526	 * Note that this race is relieved by reserving one more entry
2527	 * in tx ring than it is necessary (see original non-SG driver).
2528	 * However, with SG we need to reserve 2*MAX_SKB_FRAGS+1, which
2529	 * is already overkill.
2530	 *
2531	 * Alternative is to return with 1 not throttling queue. In this
2532	 * case loop becomes longer, no more useful effects.
2533	 */
2534	if (time_before(jiffies, maxjiff)) {
2535		barrier();
2536		cpu_relax();
2537		goto restart;
2538	}
2539
2540	/* The ring is stuck full. */
2541	printk(KERN_WARNING "%s: Transmit ring stuck full\n", dev->name);
2542	return NETDEV_TX_BUSY;
2543}
2544
2545
2546static int ace_change_mtu(struct net_device *dev, int new_mtu)
2547{
2548	struct ace_private *ap = netdev_priv(dev);
2549	struct ace_regs __iomem *regs = ap->regs;
2550
2551	if (new_mtu > ACE_JUMBO_MTU)
2552		return -EINVAL;
2553
2554	writel(new_mtu + ETH_HLEN + 4, &regs->IfMtu);
2555	dev->mtu = new_mtu;
2556
2557	if (new_mtu > ACE_STD_MTU) {
2558		if (!(ap->jumbo)) {
2559			printk(KERN_INFO "%s: Enabling Jumbo frame "
2560			       "support\n", dev->name);
2561			ap->jumbo = 1;
2562			if (!test_and_set_bit(0, &ap->jumbo_refill_busy))
2563				ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE);
2564			ace_set_rxtx_parms(dev, 1);
2565		}
2566	} else {
2567		while (test_and_set_bit(0, &ap->jumbo_refill_busy));
2568		ace_sync_irq(dev->irq);
2569		ace_set_rxtx_parms(dev, 0);
2570		if (ap->jumbo) {
2571			struct cmd cmd;
2572
2573			cmd.evt = C_RESET_JUMBO_RNG;
2574			cmd.code = 0;
2575			cmd.idx = 0;
2576			ace_issue_cmd(regs, &cmd);
2577		}
2578	}
2579
2580	return 0;
2581}
2582
2583static int ace_get_settings(struct net_device *dev, struct ethtool_cmd *ecmd)
2584{
2585	struct ace_private *ap = netdev_priv(dev);
2586	struct ace_regs __iomem *regs = ap->regs;
2587	u32 link;
2588
2589	memset(ecmd, 0, sizeof(struct ethtool_cmd));
2590	ecmd->supported =
2591		(SUPPORTED_10baseT_Half | SUPPORTED_10baseT_Full |
2592		 SUPPORTED_100baseT_Half | SUPPORTED_100baseT_Full |
2593		 SUPPORTED_1000baseT_Half | SUPPORTED_1000baseT_Full |
2594		 SUPPORTED_Autoneg | SUPPORTED_FIBRE);
2595
2596	ecmd->port = PORT_FIBRE;
2597	ecmd->transceiver = XCVR_INTERNAL;
2598
2599	link = readl(&regs->GigLnkState);
2600	if (link & LNK_1000MB)
2601		ethtool_cmd_speed_set(ecmd, SPEED_1000);
2602	else {
2603		link = readl(&regs->FastLnkState);
2604		if (link & LNK_100MB)
2605			ethtool_cmd_speed_set(ecmd, SPEED_100);
2606		else if (link & LNK_10MB)
2607			ethtool_cmd_speed_set(ecmd, SPEED_10);
2608		else
2609			ethtool_cmd_speed_set(ecmd, 0);
2610	}
2611	if (link & LNK_FULL_DUPLEX)
2612		ecmd->duplex = DUPLEX_FULL;
2613	else
2614		ecmd->duplex = DUPLEX_HALF;
2615
2616	if (link & LNK_NEGOTIATE)
2617		ecmd->autoneg = AUTONEG_ENABLE;
2618	else
2619		ecmd->autoneg = AUTONEG_DISABLE;
2620
2621#if 0
2622	/*
2623	 * Current struct ethtool_cmd is insufficient
2624	 */
2625	ecmd->trace = readl(&regs->TuneTrace);
2626
2627	ecmd->txcoal = readl(&regs->TuneTxCoalTicks);
2628	ecmd->rxcoal = readl(&regs->TuneRxCoalTicks);
2629#endif
2630	ecmd->maxtxpkt = readl(&regs->TuneMaxTxDesc);
2631	ecmd->maxrxpkt = readl(&regs->TuneMaxRxDesc);
2632
2633	return 0;
2634}
2635
2636static int ace_set_settings(struct net_device *dev, struct ethtool_cmd *ecmd)
2637{
2638	struct ace_private *ap = netdev_priv(dev);
2639	struct ace_regs __iomem *regs = ap->regs;
2640	u32 link, speed;
2641
2642	link = readl(&regs->GigLnkState);
2643	if (link & LNK_1000MB)
2644		speed = SPEED_1000;
2645	else {
2646		link = readl(&regs->FastLnkState);
2647		if (link & LNK_100MB)
2648			speed = SPEED_100;
2649		else if (link & LNK_10MB)
2650			speed = SPEED_10;
2651		else
2652			speed = SPEED_100;
2653	}
2654
2655	link = LNK_ENABLE | LNK_1000MB | LNK_100MB | LNK_10MB |
2656		LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL;
2657	if (!ACE_IS_TIGON_I(ap))
2658		link |= LNK_TX_FLOW_CTL_Y;
2659	if (ecmd->autoneg == AUTONEG_ENABLE)
2660		link |= LNK_NEGOTIATE;
2661	if (ethtool_cmd_speed(ecmd) != speed) {
2662		link &= ~(LNK_1000MB | LNK_100MB | LNK_10MB);
2663		switch (ethtool_cmd_speed(ecmd)) {
2664		case SPEED_1000:
2665			link |= LNK_1000MB;
2666			break;
2667		case SPEED_100:
2668			link |= LNK_100MB;
2669			break;
2670		case SPEED_10:
2671			link |= LNK_10MB;
2672			break;
2673		}
2674	}
2675
2676	if (ecmd->duplex == DUPLEX_FULL)
2677		link |= LNK_FULL_DUPLEX;
2678
2679	if (link != ap->link) {
2680		struct cmd cmd;
2681		printk(KERN_INFO "%s: Renegotiating link state\n",
2682		       dev->name);
2683
2684		ap->link = link;
2685		writel(link, &regs->TuneLink);
2686		if (!ACE_IS_TIGON_I(ap))
2687			writel(link, &regs->TuneFastLink);
2688		wmb();
2689
2690		cmd.evt = C_LNK_NEGOTIATION;
2691		cmd.code = 0;
2692		cmd.idx = 0;
2693		ace_issue_cmd(regs, &cmd);
2694	}
2695	return 0;
2696}
2697
2698static void ace_get_drvinfo(struct net_device *dev,
2699			    struct ethtool_drvinfo *info)
2700{
2701	struct ace_private *ap = netdev_priv(dev);
2702
2703	strlcpy(info->driver, "acenic", sizeof(info->driver));
2704	snprintf(info->version, sizeof(info->version), "%i.%i.%i",
2705		 ap->firmware_major, ap->firmware_minor,
2706		 ap->firmware_fix);
2707
2708	if (ap->pdev)
2709		strlcpy(info->bus_info, pci_name(ap->pdev),
2710			sizeof(info->bus_info));
2711
2712}
2713
2714/*
2715 * Set the hardware MAC address.
2716 */
2717static int ace_set_mac_addr(struct net_device *dev, void *p)
2718{
2719	struct ace_private *ap = netdev_priv(dev);
2720	struct ace_regs __iomem *regs = ap->regs;
2721	struct sockaddr *addr=p;
2722	u8 *da;
2723	struct cmd cmd;
2724
2725	if(netif_running(dev))
2726		return -EBUSY;
2727
2728	memcpy(dev->dev_addr, addr->sa_data,dev->addr_len);
2729
2730	da = (u8 *)dev->dev_addr;
2731
2732	writel(da[0] << 8 | da[1], &regs->MacAddrHi);
2733	writel((da[2] << 24) | (da[3] << 16) | (da[4] << 8) | da[5],
2734	       &regs->MacAddrLo);
2735
2736	cmd.evt = C_SET_MAC_ADDR;
2737	cmd.code = 0;
2738	cmd.idx = 0;
2739	ace_issue_cmd(regs, &cmd);
2740
2741	return 0;
2742}
2743
2744
2745static void ace_set_multicast_list(struct net_device *dev)
2746{
2747	struct ace_private *ap = netdev_priv(dev);
2748	struct ace_regs __iomem *regs = ap->regs;
2749	struct cmd cmd;
2750
2751	if ((dev->flags & IFF_ALLMULTI) && !(ap->mcast_all)) {
2752		cmd.evt = C_SET_MULTICAST_MODE;
2753		cmd.code = C_C_MCAST_ENABLE;
2754		cmd.idx = 0;
2755		ace_issue_cmd(regs, &cmd);
2756		ap->mcast_all = 1;
2757	} else if (ap->mcast_all) {
2758		cmd.evt = C_SET_MULTICAST_MODE;
2759		cmd.code = C_C_MCAST_DISABLE;
2760		cmd.idx = 0;
2761		ace_issue_cmd(regs, &cmd);
2762		ap->mcast_all = 0;
2763	}
2764
2765	if ((dev->flags & IFF_PROMISC) && !(ap->promisc)) {
2766		cmd.evt = C_SET_PROMISC_MODE;
2767		cmd.code = C_C_PROMISC_ENABLE;
2768		cmd.idx = 0;
2769		ace_issue_cmd(regs, &cmd);
2770		ap->promisc = 1;
2771	}else if (!(dev->flags & IFF_PROMISC) && (ap->promisc)) {
2772		cmd.evt = C_SET_PROMISC_MODE;
2773		cmd.code = C_C_PROMISC_DISABLE;
2774		cmd.idx = 0;
2775		ace_issue_cmd(regs, &cmd);
2776		ap->promisc = 0;
2777	}
2778
2779	/*
2780	 * For the time being multicast relies on the upper layers
2781	 * filtering it properly. The Firmware does not allow one to
2782	 * set the entire multicast list at a time and keeping track of
2783	 * it here is going to be messy.
2784	 */
2785	if (!netdev_mc_empty(dev) && !ap->mcast_all) {
2786		cmd.evt = C_SET_MULTICAST_MODE;
2787		cmd.code = C_C_MCAST_ENABLE;
2788		cmd.idx = 0;
2789		ace_issue_cmd(regs, &cmd);
2790	}else if (!ap->mcast_all) {
2791		cmd.evt = C_SET_MULTICAST_MODE;
2792		cmd.code = C_C_MCAST_DISABLE;
2793		cmd.idx = 0;
2794		ace_issue_cmd(regs, &cmd);
2795	}
2796}
2797
2798
2799static struct net_device_stats *ace_get_stats(struct net_device *dev)
2800{
2801	struct ace_private *ap = netdev_priv(dev);
2802	struct ace_mac_stats __iomem *mac_stats =
2803		(struct ace_mac_stats __iomem *)ap->regs->Stats;
2804
2805	dev->stats.rx_missed_errors = readl(&mac_stats->drop_space);
2806	dev->stats.multicast = readl(&mac_stats->kept_mc);
2807	dev->stats.collisions = readl(&mac_stats->coll);
2808
2809	return &dev->stats;
2810}
2811
2812
2813static void ace_copy(struct ace_regs __iomem *regs, const __be32 *src,
2814		     u32 dest, int size)
2815{
2816	void __iomem *tdest;
2817	short tsize, i;
2818
2819	if (size <= 0)
2820		return;
2821
2822	while (size > 0) {
2823		tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2824			    min_t(u32, size, ACE_WINDOW_SIZE));
2825		tdest = (void __iomem *) &regs->Window +
2826			(dest & (ACE_WINDOW_SIZE - 1));
2827		writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2828		for (i = 0; i < (tsize / 4); i++) {
2829			/* Firmware is big-endian */
2830			writel(be32_to_cpup(src), tdest);
2831			src++;
2832			tdest += 4;
2833			dest += 4;
2834			size -= 4;
2835		}
2836	}
2837}
2838
2839
2840static void ace_clear(struct ace_regs __iomem *regs, u32 dest, int size)
2841{
2842	void __iomem *tdest;
2843	short tsize = 0, i;
2844
2845	if (size <= 0)
2846		return;
2847
2848	while (size > 0) {
2849		tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2850				min_t(u32, size, ACE_WINDOW_SIZE));
2851		tdest = (void __iomem *) &regs->Window +
2852			(dest & (ACE_WINDOW_SIZE - 1));
2853		writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2854
2855		for (i = 0; i < (tsize / 4); i++) {
2856			writel(0, tdest + i*4);
2857		}
2858
2859		dest += tsize;
2860		size -= tsize;
2861	}
2862}
2863
2864
2865/*
2866 * Download the firmware into the SRAM on the NIC
2867 *
2868 * This operation requires the NIC to be halted and is performed with
2869 * interrupts disabled and with the spinlock hold.
2870 */
2871static int ace_load_firmware(struct net_device *dev)
2872{
2873	const struct firmware *fw;
2874	const char *fw_name = "acenic/tg2.bin";
2875	struct ace_private *ap = netdev_priv(dev);
2876	struct ace_regs __iomem *regs = ap->regs;
2877	const __be32 *fw_data;
2878	u32 load_addr;
2879	int ret;
2880
2881	if (!(readl(&regs->CpuCtrl) & CPU_HALTED)) {
2882		printk(KERN_ERR "%s: trying to download firmware while the "
2883		       "CPU is running!\n", ap->name);
2884		return -EFAULT;
2885	}
2886
2887	if (ACE_IS_TIGON_I(ap))
2888		fw_name = "acenic/tg1.bin";
2889
2890	ret = request_firmware(&fw, fw_name, &ap->pdev->dev);
2891	if (ret) {
2892		printk(KERN_ERR "%s: Failed to load firmware \"%s\"\n",
2893		       ap->name, fw_name);
2894		return ret;
2895	}
2896
2897	fw_data = (void *)fw->data;
2898
2899	/* Firmware blob starts with version numbers, followed by
2900	   load and start address. Remainder is the blob to be loaded
2901	   contiguously from load address. We don't bother to represent
2902	   the BSS/SBSS sections any more, since we were clearing the
2903	   whole thing anyway. */
2904	ap->firmware_major = fw->data[0];
2905	ap->firmware_minor = fw->data[1];
2906	ap->firmware_fix = fw->data[2];
2907
2908	ap->firmware_start = be32_to_cpu(fw_data[1]);
2909	if (ap->firmware_start < 0x4000 || ap->firmware_start >= 0x80000) {
2910		printk(KERN_ERR "%s: bogus load address %08x in \"%s\"\n",
2911		       ap->name, ap->firmware_start, fw_name);
2912		ret = -EINVAL;
2913		goto out;
2914	}
2915
2916	load_addr = be32_to_cpu(fw_data[2]);
2917	if (load_addr < 0x4000 || load_addr >= 0x80000) {
2918		printk(KERN_ERR "%s: bogus load address %08x in \"%s\"\n",
2919		       ap->name, load_addr, fw_name);
2920		ret = -EINVAL;
2921		goto out;
2922	}
2923
2924	/*
2925	 * Do not try to clear more than 512KiB or we end up seeing
2926	 * funny things on NICs with only 512KiB SRAM
2927	 */
2928	ace_clear(regs, 0x2000, 0x80000-0x2000);
2929	ace_copy(regs, &fw_data[3], load_addr, fw->size-12);
2930 out:
2931	release_firmware(fw);
2932	return ret;
2933}
2934
2935
2936/*
2937 * The eeprom on the AceNIC is an Atmel i2c EEPROM.
2938 *
2939 * Accessing the EEPROM is `interesting' to say the least - don't read
2940 * this code right after dinner.
2941 *
2942 * This is all about black magic and bit-banging the device .... I
2943 * wonder in what hospital they have put the guy who designed the i2c
2944 * specs.
2945 *
2946 * Oh yes, this is only the beginning!
2947 *
2948 * Thanks to Stevarino Webinski for helping tracking down the bugs in the
2949 * code i2c readout code by beta testing all my hacks.
2950 */
2951static void eeprom_start(struct ace_regs __iomem *regs)
2952{
2953	u32 local;
2954
2955	readl(&regs->LocalCtrl);
2956	udelay(ACE_SHORT_DELAY);
2957	local = readl(&regs->LocalCtrl);
2958	local |= EEPROM_DATA_OUT | EEPROM_WRITE_ENABLE;
2959	writel(local, &regs->LocalCtrl);
2960	readl(&regs->LocalCtrl);
2961	mb();
2962	udelay(ACE_SHORT_DELAY);
2963	local |= EEPROM_CLK_OUT;
2964	writel(local, &regs->LocalCtrl);
2965	readl(&regs->LocalCtrl);
2966	mb();
2967	udelay(ACE_SHORT_DELAY);
2968	local &= ~EEPROM_DATA_OUT;
2969	writel(local, &regs->LocalCtrl);
2970	readl(&regs->LocalCtrl);
2971	mb();
2972	udelay(ACE_SHORT_DELAY);
2973	local &= ~EEPROM_CLK_OUT;
2974	writel(local, &regs->LocalCtrl);
2975	readl(&regs->LocalCtrl);
2976	mb();
2977}
2978
2979
2980static void eeprom_prep(struct ace_regs __iomem *regs, u8 magic)
2981{
2982	short i;
2983	u32 local;
2984
2985	udelay(ACE_SHORT_DELAY);
2986	local = readl(&regs->LocalCtrl);
2987	local &= ~EEPROM_DATA_OUT;
2988	local |= EEPROM_WRITE_ENABLE;
2989	writel(local, &regs->LocalCtrl);
2990	readl(&regs->LocalCtrl);
2991	mb();
2992
2993	for (i = 0; i < 8; i++, magic <<= 1) {
2994		udelay(ACE_SHORT_DELAY);
2995		if (magic & 0x80)
2996			local |= EEPROM_DATA_OUT;
2997		else
2998			local &= ~EEPROM_DATA_OUT;
2999		writel(local, &regs->LocalCtrl);
3000		readl(&regs->LocalCtrl);
3001		mb();
3002
3003		udelay(ACE_SHORT_DELAY);
3004		local |= EEPROM_CLK_OUT;
3005		writel(local, &regs->LocalCtrl);
3006		readl(&regs->LocalCtrl);
3007		mb();
3008		udelay(ACE_SHORT_DELAY);
3009		local &= ~(EEPROM_CLK_OUT | EEPROM_DATA_OUT);
3010		writel(local, &regs->LocalCtrl);
3011		readl(&regs->LocalCtrl);
3012		mb();
3013	}
3014}
3015
3016
3017static int eeprom_check_ack(struct ace_regs __iomem *regs)
3018{
3019	int state;
3020	u32 local;
3021
3022	local = readl(&regs->LocalCtrl);
3023	local &= ~EEPROM_WRITE_ENABLE;
3024	writel(local, &regs->LocalCtrl);
3025	readl(&regs->LocalCtrl);
3026	mb();
3027	udelay(ACE_LONG_DELAY);
3028	local |= EEPROM_CLK_OUT;
3029	writel(local, &regs->LocalCtrl);
3030	readl(&regs->LocalCtrl);
3031	mb();
3032	udelay(ACE_SHORT_DELAY);
3033	/* sample data in middle of high clk */
3034	state = (readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0;
3035	udelay(ACE_SHORT_DELAY);
3036	mb();
3037	writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3038	readl(&regs->LocalCtrl);
3039	mb();
3040
3041	return state;
3042}
3043
3044
3045static void eeprom_stop(struct ace_regs __iomem *regs)
3046{
3047	u32 local;
3048
3049	udelay(ACE_SHORT_DELAY);
3050	local = readl(&regs->LocalCtrl);
3051	local |= EEPROM_WRITE_ENABLE;
3052	writel(local, &regs->LocalCtrl);
3053	readl(&regs->LocalCtrl);
3054	mb();
3055	udelay(ACE_SHORT_DELAY);
3056	local &= ~EEPROM_DATA_OUT;
3057	writel(local, &regs->LocalCtrl);
3058	readl(&regs->LocalCtrl);
3059	mb();
3060	udelay(ACE_SHORT_DELAY);
3061	local |= EEPROM_CLK_OUT;
3062	writel(local, &regs->LocalCtrl);
3063	readl(&regs->LocalCtrl);
3064	mb();
3065	udelay(ACE_SHORT_DELAY);
3066	local |= EEPROM_DATA_OUT;
3067	writel(local, &regs->LocalCtrl);
3068	readl(&regs->LocalCtrl);
3069	mb();
3070	udelay(ACE_LONG_DELAY);
3071	local &= ~EEPROM_CLK_OUT;
3072	writel(local, &regs->LocalCtrl);
3073	mb();
3074}
3075
3076
3077/*
3078 * Read a whole byte from the EEPROM.
3079 */
3080static int read_eeprom_byte(struct net_device *dev, unsigned long offset)
3081{
3082	struct ace_private *ap = netdev_priv(dev);
3083	struct ace_regs __iomem *regs = ap->regs;
3084	unsigned long flags;
3085	u32 local;
3086	int result = 0;
3087	short i;
3088
3089	/*
3090	 * Don't take interrupts on this CPU will bit banging
3091	 * the %#%#@$ I2C device
3092	 */
3093	local_irq_save(flags);
3094
3095	eeprom_start(regs);
3096
3097	eeprom_prep(regs, EEPROM_WRITE_SELECT);
3098	if (eeprom_check_ack(regs)) {
3099		local_irq_restore(flags);
3100		printk(KERN_ERR "%s: Unable to sync eeprom\n", ap->name);
3101		result = -EIO;
3102		goto eeprom_read_error;
3103	}
3104
3105	eeprom_prep(regs, (offset >> 8) & 0xff);
3106	if (eeprom_check_ack(regs)) {
3107		local_irq_restore(flags);
3108		printk(KERN_ERR "%s: Unable to set address byte 0\n",
3109		       ap->name);
3110		result = -EIO;
3111		goto eeprom_read_error;
3112	}
3113
3114	eeprom_prep(regs, offset & 0xff);
3115	if (eeprom_check_ack(regs)) {
3116		local_irq_restore(flags);
3117		printk(KERN_ERR "%s: Unable to set address byte 1\n",
3118		       ap->name);
3119		result = -EIO;
3120		goto eeprom_read_error;
3121	}
3122
3123	eeprom_start(regs);
3124	eeprom_prep(regs, EEPROM_READ_SELECT);
3125	if (eeprom_check_ack(regs)) {
3126		local_irq_restore(flags);
3127		printk(KERN_ERR "%s: Unable to set READ_SELECT\n",
3128		       ap->name);
3129		result = -EIO;
3130		goto eeprom_read_error;
3131	}
3132
3133	for (i = 0; i < 8; i++) {
3134		local = readl(&regs->LocalCtrl);
3135		local &= ~EEPROM_WRITE_ENABLE;
3136		writel(local, &regs->LocalCtrl);
3137		readl(&regs->LocalCtrl);
3138		udelay(ACE_LONG_DELAY);
3139		mb();
3140		local |= EEPROM_CLK_OUT;
3141		writel(local, &regs->LocalCtrl);
3142		readl(&regs->LocalCtrl);
3143		mb();
3144		udelay(ACE_SHORT_DELAY);
3145		/* sample data mid high clk */
3146		result = (result << 1) |
3147			((readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0);
3148		udelay(ACE_SHORT_DELAY);
3149		mb();
3150		local = readl(&regs->LocalCtrl);
3151		local &= ~EEPROM_CLK_OUT;
3152		writel(local, &regs->LocalCtrl);
3153		readl(&regs->LocalCtrl);
3154		udelay(ACE_SHORT_DELAY);
3155		mb();
3156		if (i == 7) {
3157			local |= EEPROM_WRITE_ENABLE;
3158			writel(local, &regs->LocalCtrl);
3159			readl(&regs->LocalCtrl);
3160			mb();
3161			udelay(ACE_SHORT_DELAY);
3162		}
3163	}
3164
3165	local |= EEPROM_DATA_OUT;
3166	writel(local, &regs->LocalCtrl);
3167	readl(&regs->LocalCtrl);
3168	mb();
3169	udelay(ACE_SHORT_DELAY);
3170	writel(readl(&regs->LocalCtrl) | EEPROM_CLK_OUT, &regs->LocalCtrl);
3171	readl(&regs->LocalCtrl);
3172	udelay(ACE_LONG_DELAY);
3173	writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3174	readl(&regs->LocalCtrl);
3175	mb();
3176	udelay(ACE_SHORT_DELAY);
3177	eeprom_stop(regs);
3178
3179	local_irq_restore(flags);
3180 out:
3181	return result;
3182
3183 eeprom_read_error:
3184	printk(KERN_ERR "%s: Unable to read eeprom byte 0x%02lx\n",
3185	       ap->name, offset);
3186	goto out;
3187}
3188
3189module_pci_driver(acenic_pci_driver);
3190