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 <linux/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 
134 static 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 };
157 MODULE_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 
395 static int link_state[ACE_MAX_MOD_PARMS];
396 static int trace[ACE_MAX_MOD_PARMS];
397 static int tx_coal_tick[ACE_MAX_MOD_PARMS];
398 static int rx_coal_tick[ACE_MAX_MOD_PARMS];
399 static int max_tx_desc[ACE_MAX_MOD_PARMS];
400 static int max_rx_desc[ACE_MAX_MOD_PARMS];
401 static int tx_ratio[ACE_MAX_MOD_PARMS];
402 static int dis_pci_mem_inval[ACE_MAX_MOD_PARMS] = {1, 1, 1, 1, 1, 1, 1, 1};
403 
404 MODULE_AUTHOR("Jes Sorensen <jes@trained-monkey.org>");
405 MODULE_LICENSE("GPL");
406 MODULE_DESCRIPTION("AceNIC/3C985/GA620 Gigabit Ethernet driver");
407 #ifndef CONFIG_ACENIC_OMIT_TIGON_I
408 MODULE_FIRMWARE("acenic/tg1.bin");
409 #endif
410 MODULE_FIRMWARE("acenic/tg2.bin");
411 
412 module_param_array_named(link, link_state, int, NULL, 0);
413 module_param_array(trace, int, NULL, 0);
414 module_param_array(tx_coal_tick, int, NULL, 0);
415 module_param_array(max_tx_desc, int, NULL, 0);
416 module_param_array(rx_coal_tick, int, NULL, 0);
417 module_param_array(max_rx_desc, int, NULL, 0);
418 module_param_array(tx_ratio, int, NULL, 0);
419 MODULE_PARM_DESC(link, "AceNIC/3C985/NetGear link state");
420 MODULE_PARM_DESC(trace, "AceNIC/3C985/NetGear firmware trace level");
421 MODULE_PARM_DESC(tx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first tx descriptor arrives");
422 MODULE_PARM_DESC(max_tx_desc, "AceNIC/3C985/GA620 max number of transmit descriptors to wait");
423 MODULE_PARM_DESC(rx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first rx descriptor arrives");
424 MODULE_PARM_DESC(max_rx_desc, "AceNIC/3C985/GA620 max number of receive descriptors to wait");
425 MODULE_PARM_DESC(tx_ratio, "AceNIC/3C985/GA620 ratio of NIC memory used for TX/RX descriptors (range 0-63)");
426 
427 
428 static 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 
432 static int ace_get_link_ksettings(struct net_device *,
433 				  struct ethtool_link_ksettings *);
434 static int ace_set_link_ksettings(struct net_device *,
435 				  const struct ethtool_link_ksettings *);
436 static void ace_get_drvinfo(struct net_device *, struct ethtool_drvinfo *);
437 
438 static const struct ethtool_ops ace_ethtool_ops = {
439 	.get_drvinfo = ace_get_drvinfo,
440 	.get_link_ksettings = ace_get_link_ksettings,
441 	.set_link_ksettings = ace_set_link_ksettings,
442 };
443 
444 static void ace_watchdog(struct net_device *dev);
445 
446 static const struct net_device_ops ace_netdev_ops = {
447 	.ndo_open		= ace_open,
448 	.ndo_stop		= ace_close,
449 	.ndo_tx_timeout		= ace_watchdog,
450 	.ndo_get_stats		= ace_get_stats,
451 	.ndo_start_xmit		= ace_start_xmit,
452 	.ndo_set_rx_mode	= ace_set_multicast_list,
453 	.ndo_validate_addr	= eth_validate_addr,
454 	.ndo_set_mac_address	= ace_set_mac_addr,
455 	.ndo_change_mtu		= ace_change_mtu,
456 };
457 
458 static int acenic_probe_one(struct pci_dev *pdev,
459 			    const struct pci_device_id *id)
460 {
461 	struct net_device *dev;
462 	struct ace_private *ap;
463 	static int boards_found;
464 
465 	dev = alloc_etherdev(sizeof(struct ace_private));
466 	if (dev == NULL)
467 		return -ENOMEM;
468 
469 	SET_NETDEV_DEV(dev, &pdev->dev);
470 
471 	ap = netdev_priv(dev);
472 	ap->pdev = pdev;
473 	ap->name = pci_name(pdev);
474 
475 	dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
476 	dev->features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX;
477 
478 	dev->watchdog_timeo = 5*HZ;
479 	dev->min_mtu = 0;
480 	dev->max_mtu = ACE_JUMBO_MTU;
481 
482 	dev->netdev_ops = &ace_netdev_ops;
483 	dev->ethtool_ops = &ace_ethtool_ops;
484 
485 	/* we only display this string ONCE */
486 	if (!boards_found)
487 		printk(version);
488 
489 	if (pci_enable_device(pdev))
490 		goto fail_free_netdev;
491 
492 	/*
493 	 * Enable master mode before we start playing with the
494 	 * pci_command word since pci_set_master() will modify
495 	 * it.
496 	 */
497 	pci_set_master(pdev);
498 
499 	pci_read_config_word(pdev, PCI_COMMAND, &ap->pci_command);
500 
501 	/* OpenFirmware on Mac's does not set this - DOH.. */
502 	if (!(ap->pci_command & PCI_COMMAND_MEMORY)) {
503 		printk(KERN_INFO "%s: Enabling PCI Memory Mapped "
504 		       "access - was not enabled by BIOS/Firmware\n",
505 		       ap->name);
506 		ap->pci_command = ap->pci_command | PCI_COMMAND_MEMORY;
507 		pci_write_config_word(ap->pdev, PCI_COMMAND,
508 				      ap->pci_command);
509 		wmb();
510 	}
511 
512 	pci_read_config_byte(pdev, PCI_LATENCY_TIMER, &ap->pci_latency);
513 	if (ap->pci_latency <= 0x40) {
514 		ap->pci_latency = 0x40;
515 		pci_write_config_byte(pdev, PCI_LATENCY_TIMER, ap->pci_latency);
516 	}
517 
518 	/*
519 	 * Remap the regs into kernel space - this is abuse of
520 	 * dev->base_addr since it was means for I/O port
521 	 * addresses but who gives a damn.
522 	 */
523 	dev->base_addr = pci_resource_start(pdev, 0);
524 	ap->regs = ioremap(dev->base_addr, 0x4000);
525 	if (!ap->regs) {
526 		printk(KERN_ERR "%s:  Unable to map I/O register, "
527 		       "AceNIC %i will be disabled.\n",
528 		       ap->name, boards_found);
529 		goto fail_free_netdev;
530 	}
531 
532 	switch(pdev->vendor) {
533 	case PCI_VENDOR_ID_ALTEON:
534 		if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9100T) {
535 			printk(KERN_INFO "%s: Farallon PN9100-T ",
536 			       ap->name);
537 		} else {
538 			printk(KERN_INFO "%s: Alteon AceNIC ",
539 			       ap->name);
540 		}
541 		break;
542 	case PCI_VENDOR_ID_3COM:
543 		printk(KERN_INFO "%s: 3Com 3C985 ", ap->name);
544 		break;
545 	case PCI_VENDOR_ID_NETGEAR:
546 		printk(KERN_INFO "%s: NetGear GA620 ", ap->name);
547 		break;
548 	case PCI_VENDOR_ID_DEC:
549 		if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9000SX) {
550 			printk(KERN_INFO "%s: Farallon PN9000-SX ",
551 			       ap->name);
552 			break;
553 		}
554 		/* Fall through */
555 	case PCI_VENDOR_ID_SGI:
556 		printk(KERN_INFO "%s: SGI AceNIC ", ap->name);
557 		break;
558 	default:
559 		printk(KERN_INFO "%s: Unknown AceNIC ", ap->name);
560 		break;
561 	}
562 
563 	printk("Gigabit Ethernet at 0x%08lx, ", dev->base_addr);
564 	printk("irq %d\n", pdev->irq);
565 
566 #ifdef CONFIG_ACENIC_OMIT_TIGON_I
567 	if ((readl(&ap->regs->HostCtrl) >> 28) == 4) {
568 		printk(KERN_ERR "%s: Driver compiled without Tigon I"
569 		       " support - NIC disabled\n", dev->name);
570 		goto fail_uninit;
571 	}
572 #endif
573 
574 	if (ace_allocate_descriptors(dev))
575 		goto fail_free_netdev;
576 
577 #ifdef MODULE
578 	if (boards_found >= ACE_MAX_MOD_PARMS)
579 		ap->board_idx = BOARD_IDX_OVERFLOW;
580 	else
581 		ap->board_idx = boards_found;
582 #else
583 	ap->board_idx = BOARD_IDX_STATIC;
584 #endif
585 
586 	if (ace_init(dev))
587 		goto fail_free_netdev;
588 
589 	if (register_netdev(dev)) {
590 		printk(KERN_ERR "acenic: device registration failed\n");
591 		goto fail_uninit;
592 	}
593 	ap->name = dev->name;
594 
595 	if (ap->pci_using_dac)
596 		dev->features |= NETIF_F_HIGHDMA;
597 
598 	pci_set_drvdata(pdev, dev);
599 
600 	boards_found++;
601 	return 0;
602 
603  fail_uninit:
604 	ace_init_cleanup(dev);
605  fail_free_netdev:
606 	free_netdev(dev);
607 	return -ENODEV;
608 }
609 
610 static void acenic_remove_one(struct pci_dev *pdev)
611 {
612 	struct net_device *dev = pci_get_drvdata(pdev);
613 	struct ace_private *ap = netdev_priv(dev);
614 	struct ace_regs __iomem *regs = ap->regs;
615 	short i;
616 
617 	unregister_netdev(dev);
618 
619 	writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
620 	if (ap->version >= 2)
621 		writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
622 
623 	/*
624 	 * This clears any pending interrupts
625 	 */
626 	writel(1, &regs->Mb0Lo);
627 	readl(&regs->CpuCtrl);	/* flush */
628 
629 	/*
630 	 * Make sure no other CPUs are processing interrupts
631 	 * on the card before the buffers are being released.
632 	 * Otherwise one might experience some `interesting'
633 	 * effects.
634 	 *
635 	 * Then release the RX buffers - jumbo buffers were
636 	 * already released in ace_close().
637 	 */
638 	ace_sync_irq(dev->irq);
639 
640 	for (i = 0; i < RX_STD_RING_ENTRIES; i++) {
641 		struct sk_buff *skb = ap->skb->rx_std_skbuff[i].skb;
642 
643 		if (skb) {
644 			struct ring_info *ringp;
645 			dma_addr_t mapping;
646 
647 			ringp = &ap->skb->rx_std_skbuff[i];
648 			mapping = dma_unmap_addr(ringp, mapping);
649 			pci_unmap_page(ap->pdev, mapping,
650 				       ACE_STD_BUFSIZE,
651 				       PCI_DMA_FROMDEVICE);
652 
653 			ap->rx_std_ring[i].size = 0;
654 			ap->skb->rx_std_skbuff[i].skb = NULL;
655 			dev_kfree_skb(skb);
656 		}
657 	}
658 
659 	if (ap->version >= 2) {
660 		for (i = 0; i < RX_MINI_RING_ENTRIES; i++) {
661 			struct sk_buff *skb = ap->skb->rx_mini_skbuff[i].skb;
662 
663 			if (skb) {
664 				struct ring_info *ringp;
665 				dma_addr_t mapping;
666 
667 				ringp = &ap->skb->rx_mini_skbuff[i];
668 				mapping = dma_unmap_addr(ringp,mapping);
669 				pci_unmap_page(ap->pdev, mapping,
670 					       ACE_MINI_BUFSIZE,
671 					       PCI_DMA_FROMDEVICE);
672 
673 				ap->rx_mini_ring[i].size = 0;
674 				ap->skb->rx_mini_skbuff[i].skb = NULL;
675 				dev_kfree_skb(skb);
676 			}
677 		}
678 	}
679 
680 	for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
681 		struct sk_buff *skb = ap->skb->rx_jumbo_skbuff[i].skb;
682 		if (skb) {
683 			struct ring_info *ringp;
684 			dma_addr_t mapping;
685 
686 			ringp = &ap->skb->rx_jumbo_skbuff[i];
687 			mapping = dma_unmap_addr(ringp, mapping);
688 			pci_unmap_page(ap->pdev, mapping,
689 				       ACE_JUMBO_BUFSIZE,
690 				       PCI_DMA_FROMDEVICE);
691 
692 			ap->rx_jumbo_ring[i].size = 0;
693 			ap->skb->rx_jumbo_skbuff[i].skb = NULL;
694 			dev_kfree_skb(skb);
695 		}
696 	}
697 
698 	ace_init_cleanup(dev);
699 	free_netdev(dev);
700 }
701 
702 static struct pci_driver acenic_pci_driver = {
703 	.name		= "acenic",
704 	.id_table	= acenic_pci_tbl,
705 	.probe		= acenic_probe_one,
706 	.remove		= acenic_remove_one,
707 };
708 
709 static void ace_free_descriptors(struct net_device *dev)
710 {
711 	struct ace_private *ap = netdev_priv(dev);
712 	int size;
713 
714 	if (ap->rx_std_ring != NULL) {
715 		size = (sizeof(struct rx_desc) *
716 			(RX_STD_RING_ENTRIES +
717 			 RX_JUMBO_RING_ENTRIES +
718 			 RX_MINI_RING_ENTRIES +
719 			 RX_RETURN_RING_ENTRIES));
720 		pci_free_consistent(ap->pdev, size, ap->rx_std_ring,
721 				    ap->rx_ring_base_dma);
722 		ap->rx_std_ring = NULL;
723 		ap->rx_jumbo_ring = NULL;
724 		ap->rx_mini_ring = NULL;
725 		ap->rx_return_ring = NULL;
726 	}
727 	if (ap->evt_ring != NULL) {
728 		size = (sizeof(struct event) * EVT_RING_ENTRIES);
729 		pci_free_consistent(ap->pdev, size, ap->evt_ring,
730 				    ap->evt_ring_dma);
731 		ap->evt_ring = NULL;
732 	}
733 	if (ap->tx_ring != NULL && !ACE_IS_TIGON_I(ap)) {
734 		size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
735 		pci_free_consistent(ap->pdev, size, ap->tx_ring,
736 				    ap->tx_ring_dma);
737 	}
738 	ap->tx_ring = NULL;
739 
740 	if (ap->evt_prd != NULL) {
741 		pci_free_consistent(ap->pdev, sizeof(u32),
742 				    (void *)ap->evt_prd, ap->evt_prd_dma);
743 		ap->evt_prd = NULL;
744 	}
745 	if (ap->rx_ret_prd != NULL) {
746 		pci_free_consistent(ap->pdev, sizeof(u32),
747 				    (void *)ap->rx_ret_prd,
748 				    ap->rx_ret_prd_dma);
749 		ap->rx_ret_prd = NULL;
750 	}
751 	if (ap->tx_csm != NULL) {
752 		pci_free_consistent(ap->pdev, sizeof(u32),
753 				    (void *)ap->tx_csm, ap->tx_csm_dma);
754 		ap->tx_csm = NULL;
755 	}
756 }
757 
758 
759 static int ace_allocate_descriptors(struct net_device *dev)
760 {
761 	struct ace_private *ap = netdev_priv(dev);
762 	int size;
763 
764 	size = (sizeof(struct rx_desc) *
765 		(RX_STD_RING_ENTRIES +
766 		 RX_JUMBO_RING_ENTRIES +
767 		 RX_MINI_RING_ENTRIES +
768 		 RX_RETURN_RING_ENTRIES));
769 
770 	ap->rx_std_ring = pci_alloc_consistent(ap->pdev, size,
771 					       &ap->rx_ring_base_dma);
772 	if (ap->rx_std_ring == NULL)
773 		goto fail;
774 
775 	ap->rx_jumbo_ring = ap->rx_std_ring + RX_STD_RING_ENTRIES;
776 	ap->rx_mini_ring = ap->rx_jumbo_ring + RX_JUMBO_RING_ENTRIES;
777 	ap->rx_return_ring = ap->rx_mini_ring + RX_MINI_RING_ENTRIES;
778 
779 	size = (sizeof(struct event) * EVT_RING_ENTRIES);
780 
781 	ap->evt_ring = pci_alloc_consistent(ap->pdev, size, &ap->evt_ring_dma);
782 
783 	if (ap->evt_ring == NULL)
784 		goto fail;
785 
786 	/*
787 	 * Only allocate a host TX ring for the Tigon II, the Tigon I
788 	 * has to use PCI registers for this ;-(
789 	 */
790 	if (!ACE_IS_TIGON_I(ap)) {
791 		size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
792 
793 		ap->tx_ring = pci_alloc_consistent(ap->pdev, size,
794 						   &ap->tx_ring_dma);
795 
796 		if (ap->tx_ring == NULL)
797 			goto fail;
798 	}
799 
800 	ap->evt_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
801 					   &ap->evt_prd_dma);
802 	if (ap->evt_prd == NULL)
803 		goto fail;
804 
805 	ap->rx_ret_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
806 					      &ap->rx_ret_prd_dma);
807 	if (ap->rx_ret_prd == NULL)
808 		goto fail;
809 
810 	ap->tx_csm = pci_alloc_consistent(ap->pdev, sizeof(u32),
811 					  &ap->tx_csm_dma);
812 	if (ap->tx_csm == NULL)
813 		goto fail;
814 
815 	return 0;
816 
817 fail:
818 	/* Clean up. */
819 	ace_init_cleanup(dev);
820 	return 1;
821 }
822 
823 
824 /*
825  * Generic cleanup handling data allocated during init. Used when the
826  * module is unloaded or if an error occurs during initialization
827  */
828 static void ace_init_cleanup(struct net_device *dev)
829 {
830 	struct ace_private *ap;
831 
832 	ap = netdev_priv(dev);
833 
834 	ace_free_descriptors(dev);
835 
836 	if (ap->info)
837 		pci_free_consistent(ap->pdev, sizeof(struct ace_info),
838 				    ap->info, ap->info_dma);
839 	kfree(ap->skb);
840 	kfree(ap->trace_buf);
841 
842 	if (dev->irq)
843 		free_irq(dev->irq, dev);
844 
845 	iounmap(ap->regs);
846 }
847 
848 
849 /*
850  * Commands are considered to be slow.
851  */
852 static inline void ace_issue_cmd(struct ace_regs __iomem *regs, struct cmd *cmd)
853 {
854 	u32 idx;
855 
856 	idx = readl(&regs->CmdPrd);
857 
858 	writel(*(u32 *)(cmd), &regs->CmdRng[idx]);
859 	idx = (idx + 1) % CMD_RING_ENTRIES;
860 
861 	writel(idx, &regs->CmdPrd);
862 }
863 
864 
865 static int ace_init(struct net_device *dev)
866 {
867 	struct ace_private *ap;
868 	struct ace_regs __iomem *regs;
869 	struct ace_info *info = NULL;
870 	struct pci_dev *pdev;
871 	unsigned long myjif;
872 	u64 tmp_ptr;
873 	u32 tig_ver, mac1, mac2, tmp, pci_state;
874 	int board_idx, ecode = 0;
875 	short i;
876 	unsigned char cache_size;
877 
878 	ap = netdev_priv(dev);
879 	regs = ap->regs;
880 
881 	board_idx = ap->board_idx;
882 
883 	/*
884 	 * aman@sgi.com - its useful to do a NIC reset here to
885 	 * address the `Firmware not running' problem subsequent
886 	 * to any crashes involving the NIC
887 	 */
888 	writel(HW_RESET | (HW_RESET << 24), &regs->HostCtrl);
889 	readl(&regs->HostCtrl);		/* PCI write posting */
890 	udelay(5);
891 
892 	/*
893 	 * Don't access any other registers before this point!
894 	 */
895 #ifdef __BIG_ENDIAN
896 	/*
897 	 * This will most likely need BYTE_SWAP once we switch
898 	 * to using __raw_writel()
899 	 */
900 	writel((WORD_SWAP | CLR_INT | ((WORD_SWAP | CLR_INT) << 24)),
901 	       &regs->HostCtrl);
902 #else
903 	writel((CLR_INT | WORD_SWAP | ((CLR_INT | WORD_SWAP) << 24)),
904 	       &regs->HostCtrl);
905 #endif
906 	readl(&regs->HostCtrl);		/* PCI write posting */
907 
908 	/*
909 	 * Stop the NIC CPU and clear pending interrupts
910 	 */
911 	writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
912 	readl(&regs->CpuCtrl);		/* PCI write posting */
913 	writel(0, &regs->Mb0Lo);
914 
915 	tig_ver = readl(&regs->HostCtrl) >> 28;
916 
917 	switch(tig_ver){
918 #ifndef CONFIG_ACENIC_OMIT_TIGON_I
919 	case 4:
920 	case 5:
921 		printk(KERN_INFO "  Tigon I  (Rev. %i), Firmware: %i.%i.%i, ",
922 		       tig_ver, ap->firmware_major, ap->firmware_minor,
923 		       ap->firmware_fix);
924 		writel(0, &regs->LocalCtrl);
925 		ap->version = 1;
926 		ap->tx_ring_entries = TIGON_I_TX_RING_ENTRIES;
927 		break;
928 #endif
929 	case 6:
930 		printk(KERN_INFO "  Tigon II (Rev. %i), Firmware: %i.%i.%i, ",
931 		       tig_ver, ap->firmware_major, ap->firmware_minor,
932 		       ap->firmware_fix);
933 		writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
934 		readl(&regs->CpuBCtrl);		/* PCI write posting */
935 		/*
936 		 * The SRAM bank size does _not_ indicate the amount
937 		 * of memory on the card, it controls the _bank_ size!
938 		 * Ie. a 1MB AceNIC will have two banks of 512KB.
939 		 */
940 		writel(SRAM_BANK_512K, &regs->LocalCtrl);
941 		writel(SYNC_SRAM_TIMING, &regs->MiscCfg);
942 		ap->version = 2;
943 		ap->tx_ring_entries = MAX_TX_RING_ENTRIES;
944 		break;
945 	default:
946 		printk(KERN_WARNING "  Unsupported Tigon version detected "
947 		       "(%i)\n", tig_ver);
948 		ecode = -ENODEV;
949 		goto init_error;
950 	}
951 
952 	/*
953 	 * ModeStat _must_ be set after the SRAM settings as this change
954 	 * seems to corrupt the ModeStat and possible other registers.
955 	 * The SRAM settings survive resets and setting it to the same
956 	 * value a second time works as well. This is what caused the
957 	 * `Firmware not running' problem on the Tigon II.
958 	 */
959 #ifdef __BIG_ENDIAN
960 	writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL | ACE_BYTE_SWAP_BD |
961 	       ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
962 #else
963 	writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL |
964 	       ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
965 #endif
966 	readl(&regs->ModeStat);		/* PCI write posting */
967 
968 	mac1 = 0;
969 	for(i = 0; i < 4; i++) {
970 		int t;
971 
972 		mac1 = mac1 << 8;
973 		t = read_eeprom_byte(dev, 0x8c+i);
974 		if (t < 0) {
975 			ecode = -EIO;
976 			goto init_error;
977 		} else
978 			mac1 |= (t & 0xff);
979 	}
980 	mac2 = 0;
981 	for(i = 4; i < 8; i++) {
982 		int t;
983 
984 		mac2 = mac2 << 8;
985 		t = read_eeprom_byte(dev, 0x8c+i);
986 		if (t < 0) {
987 			ecode = -EIO;
988 			goto init_error;
989 		} else
990 			mac2 |= (t & 0xff);
991 	}
992 
993 	writel(mac1, &regs->MacAddrHi);
994 	writel(mac2, &regs->MacAddrLo);
995 
996 	dev->dev_addr[0] = (mac1 >> 8) & 0xff;
997 	dev->dev_addr[1] = mac1 & 0xff;
998 	dev->dev_addr[2] = (mac2 >> 24) & 0xff;
999 	dev->dev_addr[3] = (mac2 >> 16) & 0xff;
1000 	dev->dev_addr[4] = (mac2 >> 8) & 0xff;
1001 	dev->dev_addr[5] = mac2 & 0xff;
1002 
1003 	printk("MAC: %pM\n", dev->dev_addr);
1004 
1005 	/*
1006 	 * Looks like this is necessary to deal with on all architectures,
1007 	 * even this %$#%$# N440BX Intel based thing doesn't get it right.
1008 	 * Ie. having two NICs in the machine, one will have the cache
1009 	 * line set at boot time, the other will not.
1010 	 */
1011 	pdev = ap->pdev;
1012 	pci_read_config_byte(pdev, PCI_CACHE_LINE_SIZE, &cache_size);
1013 	cache_size <<= 2;
1014 	if (cache_size != SMP_CACHE_BYTES) {
1015 		printk(KERN_INFO "  PCI cache line size set incorrectly "
1016 		       "(%i bytes) by BIOS/FW, ", cache_size);
1017 		if (cache_size > SMP_CACHE_BYTES)
1018 			printk("expecting %i\n", SMP_CACHE_BYTES);
1019 		else {
1020 			printk("correcting to %i\n", SMP_CACHE_BYTES);
1021 			pci_write_config_byte(pdev, PCI_CACHE_LINE_SIZE,
1022 					      SMP_CACHE_BYTES >> 2);
1023 		}
1024 	}
1025 
1026 	pci_state = readl(&regs->PciState);
1027 	printk(KERN_INFO "  PCI bus width: %i bits, speed: %iMHz, "
1028 	       "latency: %i clks\n",
1029 	       	(pci_state & PCI_32BIT) ? 32 : 64,
1030 		(pci_state & PCI_66MHZ) ? 66 : 33,
1031 		ap->pci_latency);
1032 
1033 	/*
1034 	 * Set the max DMA transfer size. Seems that for most systems
1035 	 * the performance is better when no MAX parameter is
1036 	 * set. However for systems enabling PCI write and invalidate,
1037 	 * DMA writes must be set to the L1 cache line size to get
1038 	 * optimal performance.
1039 	 *
1040 	 * The default is now to turn the PCI write and invalidate off
1041 	 * - that is what Alteon does for NT.
1042 	 */
1043 	tmp = READ_CMD_MEM | WRITE_CMD_MEM;
1044 	if (ap->version >= 2) {
1045 		tmp |= (MEM_READ_MULTIPLE | (pci_state & PCI_66MHZ));
1046 		/*
1047 		 * Tuning parameters only supported for 8 cards
1048 		 */
1049 		if (board_idx == BOARD_IDX_OVERFLOW ||
1050 		    dis_pci_mem_inval[board_idx]) {
1051 			if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1052 				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1053 				pci_write_config_word(pdev, PCI_COMMAND,
1054 						      ap->pci_command);
1055 				printk(KERN_INFO "  Disabling PCI memory "
1056 				       "write and invalidate\n");
1057 			}
1058 		} else if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1059 			printk(KERN_INFO "  PCI memory write & invalidate "
1060 			       "enabled by BIOS, enabling counter measures\n");
1061 
1062 			switch(SMP_CACHE_BYTES) {
1063 			case 16:
1064 				tmp |= DMA_WRITE_MAX_16;
1065 				break;
1066 			case 32:
1067 				tmp |= DMA_WRITE_MAX_32;
1068 				break;
1069 			case 64:
1070 				tmp |= DMA_WRITE_MAX_64;
1071 				break;
1072 			case 128:
1073 				tmp |= DMA_WRITE_MAX_128;
1074 				break;
1075 			default:
1076 				printk(KERN_INFO "  Cache line size %i not "
1077 				       "supported, PCI write and invalidate "
1078 				       "disabled\n", SMP_CACHE_BYTES);
1079 				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1080 				pci_write_config_word(pdev, PCI_COMMAND,
1081 						      ap->pci_command);
1082 			}
1083 		}
1084 	}
1085 
1086 #ifdef __sparc__
1087 	/*
1088 	 * On this platform, we know what the best dma settings
1089 	 * are.  We use 64-byte maximum bursts, because if we
1090 	 * burst larger than the cache line size (or even cross
1091 	 * a 64byte boundary in a single burst) the UltraSparc
1092 	 * PCI controller will disconnect at 64-byte multiples.
1093 	 *
1094 	 * Read-multiple will be properly enabled above, and when
1095 	 * set will give the PCI controller proper hints about
1096 	 * prefetching.
1097 	 */
1098 	tmp &= ~DMA_READ_WRITE_MASK;
1099 	tmp |= DMA_READ_MAX_64;
1100 	tmp |= DMA_WRITE_MAX_64;
1101 #endif
1102 #ifdef __alpha__
1103 	tmp &= ~DMA_READ_WRITE_MASK;
1104 	tmp |= DMA_READ_MAX_128;
1105 	/*
1106 	 * All the docs say MUST NOT. Well, I did.
1107 	 * Nothing terrible happens, if we load wrong size.
1108 	 * Bit w&i still works better!
1109 	 */
1110 	tmp |= DMA_WRITE_MAX_128;
1111 #endif
1112 	writel(tmp, &regs->PciState);
1113 
1114 #if 0
1115 	/*
1116 	 * The Host PCI bus controller driver has to set FBB.
1117 	 * If all devices on that PCI bus support FBB, then the controller
1118 	 * can enable FBB support in the Host PCI Bus controller (or on
1119 	 * the PCI-PCI bridge if that applies).
1120 	 * -ggg
1121 	 */
1122 	/*
1123 	 * I have received reports from people having problems when this
1124 	 * bit is enabled.
1125 	 */
1126 	if (!(ap->pci_command & PCI_COMMAND_FAST_BACK)) {
1127 		printk(KERN_INFO "  Enabling PCI Fast Back to Back\n");
1128 		ap->pci_command |= PCI_COMMAND_FAST_BACK;
1129 		pci_write_config_word(pdev, PCI_COMMAND, ap->pci_command);
1130 	}
1131 #endif
1132 
1133 	/*
1134 	 * Configure DMA attributes.
1135 	 */
1136 	if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(64))) {
1137 		ap->pci_using_dac = 1;
1138 	} else if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(32))) {
1139 		ap->pci_using_dac = 0;
1140 	} else {
1141 		ecode = -ENODEV;
1142 		goto init_error;
1143 	}
1144 
1145 	/*
1146 	 * Initialize the generic info block and the command+event rings
1147 	 * and the control blocks for the transmit and receive rings
1148 	 * as they need to be setup once and for all.
1149 	 */
1150 	if (!(info = pci_alloc_consistent(ap->pdev, sizeof(struct ace_info),
1151 					  &ap->info_dma))) {
1152 		ecode = -EAGAIN;
1153 		goto init_error;
1154 	}
1155 	ap->info = info;
1156 
1157 	/*
1158 	 * Get the memory for the skb rings.
1159 	 */
1160 	if (!(ap->skb = kmalloc(sizeof(struct ace_skb), GFP_KERNEL))) {
1161 		ecode = -EAGAIN;
1162 		goto init_error;
1163 	}
1164 
1165 	ecode = request_irq(pdev->irq, ace_interrupt, IRQF_SHARED,
1166 			    DRV_NAME, dev);
1167 	if (ecode) {
1168 		printk(KERN_WARNING "%s: Requested IRQ %d is busy\n",
1169 		       DRV_NAME, pdev->irq);
1170 		goto init_error;
1171 	} else
1172 		dev->irq = pdev->irq;
1173 
1174 #ifdef INDEX_DEBUG
1175 	spin_lock_init(&ap->debug_lock);
1176 	ap->last_tx = ACE_TX_RING_ENTRIES(ap) - 1;
1177 	ap->last_std_rx = 0;
1178 	ap->last_mini_rx = 0;
1179 #endif
1180 
1181 	memset(ap->info, 0, sizeof(struct ace_info));
1182 	memset(ap->skb, 0, sizeof(struct ace_skb));
1183 
1184 	ecode = ace_load_firmware(dev);
1185 	if (ecode)
1186 		goto init_error;
1187 
1188 	ap->fw_running = 0;
1189 
1190 	tmp_ptr = ap->info_dma;
1191 	writel(tmp_ptr >> 32, &regs->InfoPtrHi);
1192 	writel(tmp_ptr & 0xffffffff, &regs->InfoPtrLo);
1193 
1194 	memset(ap->evt_ring, 0, EVT_RING_ENTRIES * sizeof(struct event));
1195 
1196 	set_aceaddr(&info->evt_ctrl.rngptr, ap->evt_ring_dma);
1197 	info->evt_ctrl.flags = 0;
1198 
1199 	*(ap->evt_prd) = 0;
1200 	wmb();
1201 	set_aceaddr(&info->evt_prd_ptr, ap->evt_prd_dma);
1202 	writel(0, &regs->EvtCsm);
1203 
1204 	set_aceaddr(&info->cmd_ctrl.rngptr, 0x100);
1205 	info->cmd_ctrl.flags = 0;
1206 	info->cmd_ctrl.max_len = 0;
1207 
1208 	for (i = 0; i < CMD_RING_ENTRIES; i++)
1209 		writel(0, &regs->CmdRng[i]);
1210 
1211 	writel(0, &regs->CmdPrd);
1212 	writel(0, &regs->CmdCsm);
1213 
1214 	tmp_ptr = ap->info_dma;
1215 	tmp_ptr += (unsigned long) &(((struct ace_info *)0)->s.stats);
1216 	set_aceaddr(&info->stats2_ptr, (dma_addr_t) tmp_ptr);
1217 
1218 	set_aceaddr(&info->rx_std_ctrl.rngptr, ap->rx_ring_base_dma);
1219 	info->rx_std_ctrl.max_len = ACE_STD_BUFSIZE;
1220 	info->rx_std_ctrl.flags =
1221 	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1222 
1223 	memset(ap->rx_std_ring, 0,
1224 	       RX_STD_RING_ENTRIES * sizeof(struct rx_desc));
1225 
1226 	for (i = 0; i < RX_STD_RING_ENTRIES; i++)
1227 		ap->rx_std_ring[i].flags = BD_FLG_TCP_UDP_SUM;
1228 
1229 	ap->rx_std_skbprd = 0;
1230 	atomic_set(&ap->cur_rx_bufs, 0);
1231 
1232 	set_aceaddr(&info->rx_jumbo_ctrl.rngptr,
1233 		    (ap->rx_ring_base_dma +
1234 		     (sizeof(struct rx_desc) * RX_STD_RING_ENTRIES)));
1235 	info->rx_jumbo_ctrl.max_len = 0;
1236 	info->rx_jumbo_ctrl.flags =
1237 	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1238 
1239 	memset(ap->rx_jumbo_ring, 0,
1240 	       RX_JUMBO_RING_ENTRIES * sizeof(struct rx_desc));
1241 
1242 	for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++)
1243 		ap->rx_jumbo_ring[i].flags = BD_FLG_TCP_UDP_SUM | BD_FLG_JUMBO;
1244 
1245 	ap->rx_jumbo_skbprd = 0;
1246 	atomic_set(&ap->cur_jumbo_bufs, 0);
1247 
1248 	memset(ap->rx_mini_ring, 0,
1249 	       RX_MINI_RING_ENTRIES * sizeof(struct rx_desc));
1250 
1251 	if (ap->version >= 2) {
1252 		set_aceaddr(&info->rx_mini_ctrl.rngptr,
1253 			    (ap->rx_ring_base_dma +
1254 			     (sizeof(struct rx_desc) *
1255 			      (RX_STD_RING_ENTRIES +
1256 			       RX_JUMBO_RING_ENTRIES))));
1257 		info->rx_mini_ctrl.max_len = ACE_MINI_SIZE;
1258 		info->rx_mini_ctrl.flags =
1259 		  RCB_FLG_TCP_UDP_SUM|RCB_FLG_NO_PSEUDO_HDR|RCB_FLG_VLAN_ASSIST;
1260 
1261 		for (i = 0; i < RX_MINI_RING_ENTRIES; i++)
1262 			ap->rx_mini_ring[i].flags =
1263 				BD_FLG_TCP_UDP_SUM | BD_FLG_MINI;
1264 	} else {
1265 		set_aceaddr(&info->rx_mini_ctrl.rngptr, 0);
1266 		info->rx_mini_ctrl.flags = RCB_FLG_RNG_DISABLE;
1267 		info->rx_mini_ctrl.max_len = 0;
1268 	}
1269 
1270 	ap->rx_mini_skbprd = 0;
1271 	atomic_set(&ap->cur_mini_bufs, 0);
1272 
1273 	set_aceaddr(&info->rx_return_ctrl.rngptr,
1274 		    (ap->rx_ring_base_dma +
1275 		     (sizeof(struct rx_desc) *
1276 		      (RX_STD_RING_ENTRIES +
1277 		       RX_JUMBO_RING_ENTRIES +
1278 		       RX_MINI_RING_ENTRIES))));
1279 	info->rx_return_ctrl.flags = 0;
1280 	info->rx_return_ctrl.max_len = RX_RETURN_RING_ENTRIES;
1281 
1282 	memset(ap->rx_return_ring, 0,
1283 	       RX_RETURN_RING_ENTRIES * sizeof(struct rx_desc));
1284 
1285 	set_aceaddr(&info->rx_ret_prd_ptr, ap->rx_ret_prd_dma);
1286 	*(ap->rx_ret_prd) = 0;
1287 
1288 	writel(TX_RING_BASE, &regs->WinBase);
1289 
1290 	if (ACE_IS_TIGON_I(ap)) {
1291 		ap->tx_ring = (__force struct tx_desc *) regs->Window;
1292 		for (i = 0; i < (TIGON_I_TX_RING_ENTRIES
1293 				 * sizeof(struct tx_desc)) / sizeof(u32); i++)
1294 			writel(0, (__force void __iomem *)ap->tx_ring  + i * 4);
1295 
1296 		set_aceaddr(&info->tx_ctrl.rngptr, TX_RING_BASE);
1297 	} else {
1298 		memset(ap->tx_ring, 0,
1299 		       MAX_TX_RING_ENTRIES * sizeof(struct tx_desc));
1300 
1301 		set_aceaddr(&info->tx_ctrl.rngptr, ap->tx_ring_dma);
1302 	}
1303 
1304 	info->tx_ctrl.max_len = ACE_TX_RING_ENTRIES(ap);
1305 	tmp = RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1306 
1307 	/*
1308 	 * The Tigon I does not like having the TX ring in host memory ;-(
1309 	 */
1310 	if (!ACE_IS_TIGON_I(ap))
1311 		tmp |= RCB_FLG_TX_HOST_RING;
1312 #if TX_COAL_INTS_ONLY
1313 	tmp |= RCB_FLG_COAL_INT_ONLY;
1314 #endif
1315 	info->tx_ctrl.flags = tmp;
1316 
1317 	set_aceaddr(&info->tx_csm_ptr, ap->tx_csm_dma);
1318 
1319 	/*
1320 	 * Potential item for tuning parameter
1321 	 */
1322 #if 0 /* NO */
1323 	writel(DMA_THRESH_16W, &regs->DmaReadCfg);
1324 	writel(DMA_THRESH_16W, &regs->DmaWriteCfg);
1325 #else
1326 	writel(DMA_THRESH_8W, &regs->DmaReadCfg);
1327 	writel(DMA_THRESH_8W, &regs->DmaWriteCfg);
1328 #endif
1329 
1330 	writel(0, &regs->MaskInt);
1331 	writel(1, &regs->IfIdx);
1332 #if 0
1333 	/*
1334 	 * McKinley boxes do not like us fiddling with AssistState
1335 	 * this early
1336 	 */
1337 	writel(1, &regs->AssistState);
1338 #endif
1339 
1340 	writel(DEF_STAT, &regs->TuneStatTicks);
1341 	writel(DEF_TRACE, &regs->TuneTrace);
1342 
1343 	ace_set_rxtx_parms(dev, 0);
1344 
1345 	if (board_idx == BOARD_IDX_OVERFLOW) {
1346 		printk(KERN_WARNING "%s: more than %i NICs detected, "
1347 		       "ignoring module parameters!\n",
1348 		       ap->name, ACE_MAX_MOD_PARMS);
1349 	} else if (board_idx >= 0) {
1350 		if (tx_coal_tick[board_idx])
1351 			writel(tx_coal_tick[board_idx],
1352 			       &regs->TuneTxCoalTicks);
1353 		if (max_tx_desc[board_idx])
1354 			writel(max_tx_desc[board_idx], &regs->TuneMaxTxDesc);
1355 
1356 		if (rx_coal_tick[board_idx])
1357 			writel(rx_coal_tick[board_idx],
1358 			       &regs->TuneRxCoalTicks);
1359 		if (max_rx_desc[board_idx])
1360 			writel(max_rx_desc[board_idx], &regs->TuneMaxRxDesc);
1361 
1362 		if (trace[board_idx])
1363 			writel(trace[board_idx], &regs->TuneTrace);
1364 
1365 		if ((tx_ratio[board_idx] > 0) && (tx_ratio[board_idx] < 64))
1366 			writel(tx_ratio[board_idx], &regs->TxBufRat);
1367 	}
1368 
1369 	/*
1370 	 * Default link parameters
1371 	 */
1372 	tmp = LNK_ENABLE | LNK_FULL_DUPLEX | LNK_1000MB | LNK_100MB |
1373 		LNK_10MB | LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL | LNK_NEGOTIATE;
1374 	if(ap->version >= 2)
1375 		tmp |= LNK_TX_FLOW_CTL_Y;
1376 
1377 	/*
1378 	 * Override link default parameters
1379 	 */
1380 	if ((board_idx >= 0) && link_state[board_idx]) {
1381 		int option = link_state[board_idx];
1382 
1383 		tmp = LNK_ENABLE;
1384 
1385 		if (option & 0x01) {
1386 			printk(KERN_INFO "%s: Setting half duplex link\n",
1387 			       ap->name);
1388 			tmp &= ~LNK_FULL_DUPLEX;
1389 		}
1390 		if (option & 0x02)
1391 			tmp &= ~LNK_NEGOTIATE;
1392 		if (option & 0x10)
1393 			tmp |= LNK_10MB;
1394 		if (option & 0x20)
1395 			tmp |= LNK_100MB;
1396 		if (option & 0x40)
1397 			tmp |= LNK_1000MB;
1398 		if ((option & 0x70) == 0) {
1399 			printk(KERN_WARNING "%s: No media speed specified, "
1400 			       "forcing auto negotiation\n", ap->name);
1401 			tmp |= LNK_NEGOTIATE | LNK_1000MB |
1402 				LNK_100MB | LNK_10MB;
1403 		}
1404 		if ((option & 0x100) == 0)
1405 			tmp |= LNK_NEG_FCTL;
1406 		else
1407 			printk(KERN_INFO "%s: Disabling flow control "
1408 			       "negotiation\n", ap->name);
1409 		if (option & 0x200)
1410 			tmp |= LNK_RX_FLOW_CTL_Y;
1411 		if ((option & 0x400) && (ap->version >= 2)) {
1412 			printk(KERN_INFO "%s: Enabling TX flow control\n",
1413 			       ap->name);
1414 			tmp |= LNK_TX_FLOW_CTL_Y;
1415 		}
1416 	}
1417 
1418 	ap->link = tmp;
1419 	writel(tmp, &regs->TuneLink);
1420 	if (ap->version >= 2)
1421 		writel(tmp, &regs->TuneFastLink);
1422 
1423 	writel(ap->firmware_start, &regs->Pc);
1424 
1425 	writel(0, &regs->Mb0Lo);
1426 
1427 	/*
1428 	 * Set tx_csm before we start receiving interrupts, otherwise
1429 	 * the interrupt handler might think it is supposed to process
1430 	 * tx ints before we are up and running, which may cause a null
1431 	 * pointer access in the int handler.
1432 	 */
1433 	ap->cur_rx = 0;
1434 	ap->tx_prd = *(ap->tx_csm) = ap->tx_ret_csm = 0;
1435 
1436 	wmb();
1437 	ace_set_txprd(regs, ap, 0);
1438 	writel(0, &regs->RxRetCsm);
1439 
1440 	/*
1441 	 * Enable DMA engine now.
1442 	 * If we do this sooner, Mckinley box pukes.
1443 	 * I assume it's because Tigon II DMA engine wants to check
1444 	 * *something* even before the CPU is started.
1445 	 */
1446 	writel(1, &regs->AssistState);  /* enable DMA */
1447 
1448 	/*
1449 	 * Start the NIC CPU
1450 	 */
1451 	writel(readl(&regs->CpuCtrl) & ~(CPU_HALT|CPU_TRACE), &regs->CpuCtrl);
1452 	readl(&regs->CpuCtrl);
1453 
1454 	/*
1455 	 * Wait for the firmware to spin up - max 3 seconds.
1456 	 */
1457 	myjif = jiffies + 3 * HZ;
1458 	while (time_before(jiffies, myjif) && !ap->fw_running)
1459 		cpu_relax();
1460 
1461 	if (!ap->fw_running) {
1462 		printk(KERN_ERR "%s: Firmware NOT running!\n", ap->name);
1463 
1464 		ace_dump_trace(ap);
1465 		writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
1466 		readl(&regs->CpuCtrl);
1467 
1468 		/* aman@sgi.com - account for badly behaving firmware/NIC:
1469 		 * - have observed that the NIC may continue to generate
1470 		 *   interrupts for some reason; attempt to stop it - halt
1471 		 *   second CPU for Tigon II cards, and also clear Mb0
1472 		 * - if we're a module, we'll fail to load if this was
1473 		 *   the only GbE card in the system => if the kernel does
1474 		 *   see an interrupt from the NIC, code to handle it is
1475 		 *   gone and OOps! - so free_irq also
1476 		 */
1477 		if (ap->version >= 2)
1478 			writel(readl(&regs->CpuBCtrl) | CPU_HALT,
1479 			       &regs->CpuBCtrl);
1480 		writel(0, &regs->Mb0Lo);
1481 		readl(&regs->Mb0Lo);
1482 
1483 		ecode = -EBUSY;
1484 		goto init_error;
1485 	}
1486 
1487 	/*
1488 	 * We load the ring here as there seem to be no way to tell the
1489 	 * firmware to wipe the ring without re-initializing it.
1490 	 */
1491 	if (!test_and_set_bit(0, &ap->std_refill_busy))
1492 		ace_load_std_rx_ring(dev, RX_RING_SIZE);
1493 	else
1494 		printk(KERN_ERR "%s: Someone is busy refilling the RX ring\n",
1495 		       ap->name);
1496 	if (ap->version >= 2) {
1497 		if (!test_and_set_bit(0, &ap->mini_refill_busy))
1498 			ace_load_mini_rx_ring(dev, RX_MINI_SIZE);
1499 		else
1500 			printk(KERN_ERR "%s: Someone is busy refilling "
1501 			       "the RX mini ring\n", ap->name);
1502 	}
1503 	return 0;
1504 
1505  init_error:
1506 	ace_init_cleanup(dev);
1507 	return ecode;
1508 }
1509 
1510 
1511 static void ace_set_rxtx_parms(struct net_device *dev, int jumbo)
1512 {
1513 	struct ace_private *ap = netdev_priv(dev);
1514 	struct ace_regs __iomem *regs = ap->regs;
1515 	int board_idx = ap->board_idx;
1516 
1517 	if (board_idx >= 0) {
1518 		if (!jumbo) {
1519 			if (!tx_coal_tick[board_idx])
1520 				writel(DEF_TX_COAL, &regs->TuneTxCoalTicks);
1521 			if (!max_tx_desc[board_idx])
1522 				writel(DEF_TX_MAX_DESC, &regs->TuneMaxTxDesc);
1523 			if (!rx_coal_tick[board_idx])
1524 				writel(DEF_RX_COAL, &regs->TuneRxCoalTicks);
1525 			if (!max_rx_desc[board_idx])
1526 				writel(DEF_RX_MAX_DESC, &regs->TuneMaxRxDesc);
1527 			if (!tx_ratio[board_idx])
1528 				writel(DEF_TX_RATIO, &regs->TxBufRat);
1529 		} else {
1530 			if (!tx_coal_tick[board_idx])
1531 				writel(DEF_JUMBO_TX_COAL,
1532 				       &regs->TuneTxCoalTicks);
1533 			if (!max_tx_desc[board_idx])
1534 				writel(DEF_JUMBO_TX_MAX_DESC,
1535 				       &regs->TuneMaxTxDesc);
1536 			if (!rx_coal_tick[board_idx])
1537 				writel(DEF_JUMBO_RX_COAL,
1538 				       &regs->TuneRxCoalTicks);
1539 			if (!max_rx_desc[board_idx])
1540 				writel(DEF_JUMBO_RX_MAX_DESC,
1541 				       &regs->TuneMaxRxDesc);
1542 			if (!tx_ratio[board_idx])
1543 				writel(DEF_JUMBO_TX_RATIO, &regs->TxBufRat);
1544 		}
1545 	}
1546 }
1547 
1548 
1549 static void ace_watchdog(struct net_device *data)
1550 {
1551 	struct net_device *dev = data;
1552 	struct ace_private *ap = netdev_priv(dev);
1553 	struct ace_regs __iomem *regs = ap->regs;
1554 
1555 	/*
1556 	 * We haven't received a stats update event for more than 2.5
1557 	 * seconds and there is data in the transmit queue, thus we
1558 	 * assume the card is stuck.
1559 	 */
1560 	if (*ap->tx_csm != ap->tx_ret_csm) {
1561 		printk(KERN_WARNING "%s: Transmitter is stuck, %08x\n",
1562 		       dev->name, (unsigned int)readl(&regs->HostCtrl));
1563 		/* This can happen due to ieee flow control. */
1564 	} else {
1565 		printk(KERN_DEBUG "%s: BUG... transmitter died. Kicking it.\n",
1566 		       dev->name);
1567 #if 0
1568 		netif_wake_queue(dev);
1569 #endif
1570 	}
1571 }
1572 
1573 
1574 static void ace_tasklet(unsigned long arg)
1575 {
1576 	struct net_device *dev = (struct net_device *) arg;
1577 	struct ace_private *ap = netdev_priv(dev);
1578 	int cur_size;
1579 
1580 	cur_size = atomic_read(&ap->cur_rx_bufs);
1581 	if ((cur_size < RX_LOW_STD_THRES) &&
1582 	    !test_and_set_bit(0, &ap->std_refill_busy)) {
1583 #ifdef DEBUG
1584 		printk("refilling buffers (current %i)\n", cur_size);
1585 #endif
1586 		ace_load_std_rx_ring(dev, RX_RING_SIZE - cur_size);
1587 	}
1588 
1589 	if (ap->version >= 2) {
1590 		cur_size = atomic_read(&ap->cur_mini_bufs);
1591 		if ((cur_size < RX_LOW_MINI_THRES) &&
1592 		    !test_and_set_bit(0, &ap->mini_refill_busy)) {
1593 #ifdef DEBUG
1594 			printk("refilling mini buffers (current %i)\n",
1595 			       cur_size);
1596 #endif
1597 			ace_load_mini_rx_ring(dev, RX_MINI_SIZE - cur_size);
1598 		}
1599 	}
1600 
1601 	cur_size = atomic_read(&ap->cur_jumbo_bufs);
1602 	if (ap->jumbo && (cur_size < RX_LOW_JUMBO_THRES) &&
1603 	    !test_and_set_bit(0, &ap->jumbo_refill_busy)) {
1604 #ifdef DEBUG
1605 		printk("refilling jumbo buffers (current %i)\n", cur_size);
1606 #endif
1607 		ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE - cur_size);
1608 	}
1609 	ap->tasklet_pending = 0;
1610 }
1611 
1612 
1613 /*
1614  * Copy the contents of the NIC's trace buffer to kernel memory.
1615  */
1616 static void ace_dump_trace(struct ace_private *ap)
1617 {
1618 #if 0
1619 	if (!ap->trace_buf)
1620 		if (!(ap->trace_buf = kmalloc(ACE_TRACE_SIZE, GFP_KERNEL)))
1621 		    return;
1622 #endif
1623 }
1624 
1625 
1626 /*
1627  * Load the standard rx ring.
1628  *
1629  * Loading rings is safe without holding the spin lock since this is
1630  * done only before the device is enabled, thus no interrupts are
1631  * generated and by the interrupt handler/tasklet handler.
1632  */
1633 static void ace_load_std_rx_ring(struct net_device *dev, int nr_bufs)
1634 {
1635 	struct ace_private *ap = netdev_priv(dev);
1636 	struct ace_regs __iomem *regs = ap->regs;
1637 	short i, idx;
1638 
1639 
1640 	prefetchw(&ap->cur_rx_bufs);
1641 
1642 	idx = ap->rx_std_skbprd;
1643 
1644 	for (i = 0; i < nr_bufs; i++) {
1645 		struct sk_buff *skb;
1646 		struct rx_desc *rd;
1647 		dma_addr_t mapping;
1648 
1649 		skb = netdev_alloc_skb_ip_align(dev, ACE_STD_BUFSIZE);
1650 		if (!skb)
1651 			break;
1652 
1653 		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1654 				       offset_in_page(skb->data),
1655 				       ACE_STD_BUFSIZE,
1656 				       PCI_DMA_FROMDEVICE);
1657 		ap->skb->rx_std_skbuff[idx].skb = skb;
1658 		dma_unmap_addr_set(&ap->skb->rx_std_skbuff[idx],
1659 				   mapping, mapping);
1660 
1661 		rd = &ap->rx_std_ring[idx];
1662 		set_aceaddr(&rd->addr, mapping);
1663 		rd->size = ACE_STD_BUFSIZE;
1664 		rd->idx = idx;
1665 		idx = (idx + 1) % RX_STD_RING_ENTRIES;
1666 	}
1667 
1668 	if (!i)
1669 		goto error_out;
1670 
1671 	atomic_add(i, &ap->cur_rx_bufs);
1672 	ap->rx_std_skbprd = idx;
1673 
1674 	if (ACE_IS_TIGON_I(ap)) {
1675 		struct cmd cmd;
1676 		cmd.evt = C_SET_RX_PRD_IDX;
1677 		cmd.code = 0;
1678 		cmd.idx = ap->rx_std_skbprd;
1679 		ace_issue_cmd(regs, &cmd);
1680 	} else {
1681 		writel(idx, &regs->RxStdPrd);
1682 		wmb();
1683 	}
1684 
1685  out:
1686 	clear_bit(0, &ap->std_refill_busy);
1687 	return;
1688 
1689  error_out:
1690 	printk(KERN_INFO "Out of memory when allocating "
1691 	       "standard receive buffers\n");
1692 	goto out;
1693 }
1694 
1695 
1696 static void ace_load_mini_rx_ring(struct net_device *dev, int nr_bufs)
1697 {
1698 	struct ace_private *ap = netdev_priv(dev);
1699 	struct ace_regs __iomem *regs = ap->regs;
1700 	short i, idx;
1701 
1702 	prefetchw(&ap->cur_mini_bufs);
1703 
1704 	idx = ap->rx_mini_skbprd;
1705 	for (i = 0; i < nr_bufs; i++) {
1706 		struct sk_buff *skb;
1707 		struct rx_desc *rd;
1708 		dma_addr_t mapping;
1709 
1710 		skb = netdev_alloc_skb_ip_align(dev, ACE_MINI_BUFSIZE);
1711 		if (!skb)
1712 			break;
1713 
1714 		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1715 				       offset_in_page(skb->data),
1716 				       ACE_MINI_BUFSIZE,
1717 				       PCI_DMA_FROMDEVICE);
1718 		ap->skb->rx_mini_skbuff[idx].skb = skb;
1719 		dma_unmap_addr_set(&ap->skb->rx_mini_skbuff[idx],
1720 				   mapping, mapping);
1721 
1722 		rd = &ap->rx_mini_ring[idx];
1723 		set_aceaddr(&rd->addr, mapping);
1724 		rd->size = ACE_MINI_BUFSIZE;
1725 		rd->idx = idx;
1726 		idx = (idx + 1) % RX_MINI_RING_ENTRIES;
1727 	}
1728 
1729 	if (!i)
1730 		goto error_out;
1731 
1732 	atomic_add(i, &ap->cur_mini_bufs);
1733 
1734 	ap->rx_mini_skbprd = idx;
1735 
1736 	writel(idx, &regs->RxMiniPrd);
1737 	wmb();
1738 
1739  out:
1740 	clear_bit(0, &ap->mini_refill_busy);
1741 	return;
1742  error_out:
1743 	printk(KERN_INFO "Out of memory when allocating "
1744 	       "mini receive buffers\n");
1745 	goto out;
1746 }
1747 
1748 
1749 /*
1750  * Load the jumbo rx ring, this may happen at any time if the MTU
1751  * is changed to a value > 1500.
1752  */
1753 static void ace_load_jumbo_rx_ring(struct net_device *dev, int nr_bufs)
1754 {
1755 	struct ace_private *ap = netdev_priv(dev);
1756 	struct ace_regs __iomem *regs = ap->regs;
1757 	short i, idx;
1758 
1759 	idx = ap->rx_jumbo_skbprd;
1760 
1761 	for (i = 0; i < nr_bufs; i++) {
1762 		struct sk_buff *skb;
1763 		struct rx_desc *rd;
1764 		dma_addr_t mapping;
1765 
1766 		skb = netdev_alloc_skb_ip_align(dev, ACE_JUMBO_BUFSIZE);
1767 		if (!skb)
1768 			break;
1769 
1770 		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1771 				       offset_in_page(skb->data),
1772 				       ACE_JUMBO_BUFSIZE,
1773 				       PCI_DMA_FROMDEVICE);
1774 		ap->skb->rx_jumbo_skbuff[idx].skb = skb;
1775 		dma_unmap_addr_set(&ap->skb->rx_jumbo_skbuff[idx],
1776 				   mapping, mapping);
1777 
1778 		rd = &ap->rx_jumbo_ring[idx];
1779 		set_aceaddr(&rd->addr, mapping);
1780 		rd->size = ACE_JUMBO_BUFSIZE;
1781 		rd->idx = idx;
1782 		idx = (idx + 1) % RX_JUMBO_RING_ENTRIES;
1783 	}
1784 
1785 	if (!i)
1786 		goto error_out;
1787 
1788 	atomic_add(i, &ap->cur_jumbo_bufs);
1789 	ap->rx_jumbo_skbprd = idx;
1790 
1791 	if (ACE_IS_TIGON_I(ap)) {
1792 		struct cmd cmd;
1793 		cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1794 		cmd.code = 0;
1795 		cmd.idx = ap->rx_jumbo_skbprd;
1796 		ace_issue_cmd(regs, &cmd);
1797 	} else {
1798 		writel(idx, &regs->RxJumboPrd);
1799 		wmb();
1800 	}
1801 
1802  out:
1803 	clear_bit(0, &ap->jumbo_refill_busy);
1804 	return;
1805  error_out:
1806 	if (net_ratelimit())
1807 		printk(KERN_INFO "Out of memory when allocating "
1808 		       "jumbo receive buffers\n");
1809 	goto out;
1810 }
1811 
1812 
1813 /*
1814  * All events are considered to be slow (RX/TX ints do not generate
1815  * events) and are handled here, outside the main interrupt handler,
1816  * to reduce the size of the handler.
1817  */
1818 static u32 ace_handle_event(struct net_device *dev, u32 evtcsm, u32 evtprd)
1819 {
1820 	struct ace_private *ap;
1821 
1822 	ap = netdev_priv(dev);
1823 
1824 	while (evtcsm != evtprd) {
1825 		switch (ap->evt_ring[evtcsm].evt) {
1826 		case E_FW_RUNNING:
1827 			printk(KERN_INFO "%s: Firmware up and running\n",
1828 			       ap->name);
1829 			ap->fw_running = 1;
1830 			wmb();
1831 			break;
1832 		case E_STATS_UPDATED:
1833 			break;
1834 		case E_LNK_STATE:
1835 		{
1836 			u16 code = ap->evt_ring[evtcsm].code;
1837 			switch (code) {
1838 			case E_C_LINK_UP:
1839 			{
1840 				u32 state = readl(&ap->regs->GigLnkState);
1841 				printk(KERN_WARNING "%s: Optical link UP "
1842 				       "(%s Duplex, Flow Control: %s%s)\n",
1843 				       ap->name,
1844 				       state & LNK_FULL_DUPLEX ? "Full":"Half",
1845 				       state & LNK_TX_FLOW_CTL_Y ? "TX " : "",
1846 				       state & LNK_RX_FLOW_CTL_Y ? "RX" : "");
1847 				break;
1848 			}
1849 			case E_C_LINK_DOWN:
1850 				printk(KERN_WARNING "%s: Optical link DOWN\n",
1851 				       ap->name);
1852 				break;
1853 			case E_C_LINK_10_100:
1854 				printk(KERN_WARNING "%s: 10/100BaseT link "
1855 				       "UP\n", ap->name);
1856 				break;
1857 			default:
1858 				printk(KERN_ERR "%s: Unknown optical link "
1859 				       "state %02x\n", ap->name, code);
1860 			}
1861 			break;
1862 		}
1863 		case E_ERROR:
1864 			switch(ap->evt_ring[evtcsm].code) {
1865 			case E_C_ERR_INVAL_CMD:
1866 				printk(KERN_ERR "%s: invalid command error\n",
1867 				       ap->name);
1868 				break;
1869 			case E_C_ERR_UNIMP_CMD:
1870 				printk(KERN_ERR "%s: unimplemented command "
1871 				       "error\n", ap->name);
1872 				break;
1873 			case E_C_ERR_BAD_CFG:
1874 				printk(KERN_ERR "%s: bad config error\n",
1875 				       ap->name);
1876 				break;
1877 			default:
1878 				printk(KERN_ERR "%s: unknown error %02x\n",
1879 				       ap->name, ap->evt_ring[evtcsm].code);
1880 			}
1881 			break;
1882 		case E_RESET_JUMBO_RNG:
1883 		{
1884 			int i;
1885 			for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
1886 				if (ap->skb->rx_jumbo_skbuff[i].skb) {
1887 					ap->rx_jumbo_ring[i].size = 0;
1888 					set_aceaddr(&ap->rx_jumbo_ring[i].addr, 0);
1889 					dev_kfree_skb(ap->skb->rx_jumbo_skbuff[i].skb);
1890 					ap->skb->rx_jumbo_skbuff[i].skb = NULL;
1891 				}
1892 			}
1893 
1894  			if (ACE_IS_TIGON_I(ap)) {
1895  				struct cmd cmd;
1896  				cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1897  				cmd.code = 0;
1898  				cmd.idx = 0;
1899  				ace_issue_cmd(ap->regs, &cmd);
1900  			} else {
1901  				writel(0, &((ap->regs)->RxJumboPrd));
1902  				wmb();
1903  			}
1904 
1905 			ap->jumbo = 0;
1906 			ap->rx_jumbo_skbprd = 0;
1907 			printk(KERN_INFO "%s: Jumbo ring flushed\n",
1908 			       ap->name);
1909 			clear_bit(0, &ap->jumbo_refill_busy);
1910 			break;
1911 		}
1912 		default:
1913 			printk(KERN_ERR "%s: Unhandled event 0x%02x\n",
1914 			       ap->name, ap->evt_ring[evtcsm].evt);
1915 		}
1916 		evtcsm = (evtcsm + 1) % EVT_RING_ENTRIES;
1917 	}
1918 
1919 	return evtcsm;
1920 }
1921 
1922 
1923 static void ace_rx_int(struct net_device *dev, u32 rxretprd, u32 rxretcsm)
1924 {
1925 	struct ace_private *ap = netdev_priv(dev);
1926 	u32 idx;
1927 	int mini_count = 0, std_count = 0;
1928 
1929 	idx = rxretcsm;
1930 
1931 	prefetchw(&ap->cur_rx_bufs);
1932 	prefetchw(&ap->cur_mini_bufs);
1933 
1934 	while (idx != rxretprd) {
1935 		struct ring_info *rip;
1936 		struct sk_buff *skb;
1937 		struct rx_desc *retdesc;
1938 		u32 skbidx;
1939 		int bd_flags, desc_type, mapsize;
1940 		u16 csum;
1941 
1942 
1943 		/* make sure the rx descriptor isn't read before rxretprd */
1944 		if (idx == rxretcsm)
1945 			rmb();
1946 
1947 		retdesc = &ap->rx_return_ring[idx];
1948 		skbidx = retdesc->idx;
1949 		bd_flags = retdesc->flags;
1950 		desc_type = bd_flags & (BD_FLG_JUMBO | BD_FLG_MINI);
1951 
1952 		switch(desc_type) {
1953 			/*
1954 			 * Normal frames do not have any flags set
1955 			 *
1956 			 * Mini and normal frames arrive frequently,
1957 			 * so use a local counter to avoid doing
1958 			 * atomic operations for each packet arriving.
1959 			 */
1960 		case 0:
1961 			rip = &ap->skb->rx_std_skbuff[skbidx];
1962 			mapsize = ACE_STD_BUFSIZE;
1963 			std_count++;
1964 			break;
1965 		case BD_FLG_JUMBO:
1966 			rip = &ap->skb->rx_jumbo_skbuff[skbidx];
1967 			mapsize = ACE_JUMBO_BUFSIZE;
1968 			atomic_dec(&ap->cur_jumbo_bufs);
1969 			break;
1970 		case BD_FLG_MINI:
1971 			rip = &ap->skb->rx_mini_skbuff[skbidx];
1972 			mapsize = ACE_MINI_BUFSIZE;
1973 			mini_count++;
1974 			break;
1975 		default:
1976 			printk(KERN_INFO "%s: unknown frame type (0x%02x) "
1977 			       "returned by NIC\n", dev->name,
1978 			       retdesc->flags);
1979 			goto error;
1980 		}
1981 
1982 		skb = rip->skb;
1983 		rip->skb = NULL;
1984 		pci_unmap_page(ap->pdev,
1985 			       dma_unmap_addr(rip, mapping),
1986 			       mapsize,
1987 			       PCI_DMA_FROMDEVICE);
1988 		skb_put(skb, retdesc->size);
1989 
1990 		/*
1991 		 * Fly baby, fly!
1992 		 */
1993 		csum = retdesc->tcp_udp_csum;
1994 
1995 		skb->protocol = eth_type_trans(skb, dev);
1996 
1997 		/*
1998 		 * Instead of forcing the poor tigon mips cpu to calculate
1999 		 * pseudo hdr checksum, we do this ourselves.
2000 		 */
2001 		if (bd_flags & BD_FLG_TCP_UDP_SUM) {
2002 			skb->csum = htons(csum);
2003 			skb->ip_summed = CHECKSUM_COMPLETE;
2004 		} else {
2005 			skb_checksum_none_assert(skb);
2006 		}
2007 
2008 		/* send it up */
2009 		if ((bd_flags & BD_FLG_VLAN_TAG))
2010 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), retdesc->vlan);
2011 		netif_rx(skb);
2012 
2013 		dev->stats.rx_packets++;
2014 		dev->stats.rx_bytes += retdesc->size;
2015 
2016 		idx = (idx + 1) % RX_RETURN_RING_ENTRIES;
2017 	}
2018 
2019 	atomic_sub(std_count, &ap->cur_rx_bufs);
2020 	if (!ACE_IS_TIGON_I(ap))
2021 		atomic_sub(mini_count, &ap->cur_mini_bufs);
2022 
2023  out:
2024 	/*
2025 	 * According to the documentation RxRetCsm is obsolete with
2026 	 * the 12.3.x Firmware - my Tigon I NICs seem to disagree!
2027 	 */
2028 	if (ACE_IS_TIGON_I(ap)) {
2029 		writel(idx, &ap->regs->RxRetCsm);
2030 	}
2031 	ap->cur_rx = idx;
2032 
2033 	return;
2034  error:
2035 	idx = rxretprd;
2036 	goto out;
2037 }
2038 
2039 
2040 static inline void ace_tx_int(struct net_device *dev,
2041 			      u32 txcsm, u32 idx)
2042 {
2043 	struct ace_private *ap = netdev_priv(dev);
2044 
2045 	do {
2046 		struct sk_buff *skb;
2047 		struct tx_ring_info *info;
2048 
2049 		info = ap->skb->tx_skbuff + idx;
2050 		skb = info->skb;
2051 
2052 		if (dma_unmap_len(info, maplen)) {
2053 			pci_unmap_page(ap->pdev, dma_unmap_addr(info, mapping),
2054 				       dma_unmap_len(info, maplen),
2055 				       PCI_DMA_TODEVICE);
2056 			dma_unmap_len_set(info, maplen, 0);
2057 		}
2058 
2059 		if (skb) {
2060 			dev->stats.tx_packets++;
2061 			dev->stats.tx_bytes += skb->len;
2062 			dev_consume_skb_irq(skb);
2063 			info->skb = NULL;
2064 		}
2065 
2066 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2067 	} while (idx != txcsm);
2068 
2069 	if (netif_queue_stopped(dev))
2070 		netif_wake_queue(dev);
2071 
2072 	wmb();
2073 	ap->tx_ret_csm = txcsm;
2074 
2075 	/* So... tx_ret_csm is advanced _after_ check for device wakeup.
2076 	 *
2077 	 * We could try to make it before. In this case we would get
2078 	 * the following race condition: hard_start_xmit on other cpu
2079 	 * enters after we advanced tx_ret_csm and fills space,
2080 	 * which we have just freed, so that we make illegal device wakeup.
2081 	 * There is no good way to workaround this (at entry
2082 	 * to ace_start_xmit detects this condition and prevents
2083 	 * ring corruption, but it is not a good workaround.)
2084 	 *
2085 	 * When tx_ret_csm is advanced after, we wake up device _only_
2086 	 * if we really have some space in ring (though the core doing
2087 	 * hard_start_xmit can see full ring for some period and has to
2088 	 * synchronize.) Superb.
2089 	 * BUT! We get another subtle race condition. hard_start_xmit
2090 	 * may think that ring is full between wakeup and advancing
2091 	 * tx_ret_csm and will stop device instantly! It is not so bad.
2092 	 * We are guaranteed that there is something in ring, so that
2093 	 * the next irq will resume transmission. To speedup this we could
2094 	 * mark descriptor, which closes ring with BD_FLG_COAL_NOW
2095 	 * (see ace_start_xmit).
2096 	 *
2097 	 * Well, this dilemma exists in all lock-free devices.
2098 	 * We, following scheme used in drivers by Donald Becker,
2099 	 * select the least dangerous.
2100 	 *							--ANK
2101 	 */
2102 }
2103 
2104 
2105 static irqreturn_t ace_interrupt(int irq, void *dev_id)
2106 {
2107 	struct net_device *dev = (struct net_device *)dev_id;
2108 	struct ace_private *ap = netdev_priv(dev);
2109 	struct ace_regs __iomem *regs = ap->regs;
2110 	u32 idx;
2111 	u32 txcsm, rxretcsm, rxretprd;
2112 	u32 evtcsm, evtprd;
2113 
2114 	/*
2115 	 * In case of PCI shared interrupts or spurious interrupts,
2116 	 * we want to make sure it is actually our interrupt before
2117 	 * spending any time in here.
2118 	 */
2119 	if (!(readl(&regs->HostCtrl) & IN_INT))
2120 		return IRQ_NONE;
2121 
2122 	/*
2123 	 * ACK intr now. Otherwise we will lose updates to rx_ret_prd,
2124 	 * which happened _after_ rxretprd = *ap->rx_ret_prd; but before
2125 	 * writel(0, &regs->Mb0Lo).
2126 	 *
2127 	 * "IRQ avoidance" recommended in docs applies to IRQs served
2128 	 * threads and it is wrong even for that case.
2129 	 */
2130 	writel(0, &regs->Mb0Lo);
2131 	readl(&regs->Mb0Lo);
2132 
2133 	/*
2134 	 * There is no conflict between transmit handling in
2135 	 * start_xmit and receive processing, thus there is no reason
2136 	 * to take a spin lock for RX handling. Wait until we start
2137 	 * working on the other stuff - hey we don't need a spin lock
2138 	 * anymore.
2139 	 */
2140 	rxretprd = *ap->rx_ret_prd;
2141 	rxretcsm = ap->cur_rx;
2142 
2143 	if (rxretprd != rxretcsm)
2144 		ace_rx_int(dev, rxretprd, rxretcsm);
2145 
2146 	txcsm = *ap->tx_csm;
2147 	idx = ap->tx_ret_csm;
2148 
2149 	if (txcsm != idx) {
2150 		/*
2151 		 * If each skb takes only one descriptor this check degenerates
2152 		 * to identity, because new space has just been opened.
2153 		 * But if skbs are fragmented we must check that this index
2154 		 * update releases enough of space, otherwise we just
2155 		 * wait for device to make more work.
2156 		 */
2157 		if (!tx_ring_full(ap, txcsm, ap->tx_prd))
2158 			ace_tx_int(dev, txcsm, idx);
2159 	}
2160 
2161 	evtcsm = readl(&regs->EvtCsm);
2162 	evtprd = *ap->evt_prd;
2163 
2164 	if (evtcsm != evtprd) {
2165 		evtcsm = ace_handle_event(dev, evtcsm, evtprd);
2166 		writel(evtcsm, &regs->EvtCsm);
2167 	}
2168 
2169 	/*
2170 	 * This has to go last in the interrupt handler and run with
2171 	 * the spin lock released ... what lock?
2172 	 */
2173 	if (netif_running(dev)) {
2174 		int cur_size;
2175 		int run_tasklet = 0;
2176 
2177 		cur_size = atomic_read(&ap->cur_rx_bufs);
2178 		if (cur_size < RX_LOW_STD_THRES) {
2179 			if ((cur_size < RX_PANIC_STD_THRES) &&
2180 			    !test_and_set_bit(0, &ap->std_refill_busy)) {
2181 #ifdef DEBUG
2182 				printk("low on std buffers %i\n", cur_size);
2183 #endif
2184 				ace_load_std_rx_ring(dev,
2185 						     RX_RING_SIZE - cur_size);
2186 			} else
2187 				run_tasklet = 1;
2188 		}
2189 
2190 		if (!ACE_IS_TIGON_I(ap)) {
2191 			cur_size = atomic_read(&ap->cur_mini_bufs);
2192 			if (cur_size < RX_LOW_MINI_THRES) {
2193 				if ((cur_size < RX_PANIC_MINI_THRES) &&
2194 				    !test_and_set_bit(0,
2195 						      &ap->mini_refill_busy)) {
2196 #ifdef DEBUG
2197 					printk("low on mini buffers %i\n",
2198 					       cur_size);
2199 #endif
2200 					ace_load_mini_rx_ring(dev,
2201 							      RX_MINI_SIZE - cur_size);
2202 				} else
2203 					run_tasklet = 1;
2204 			}
2205 		}
2206 
2207 		if (ap->jumbo) {
2208 			cur_size = atomic_read(&ap->cur_jumbo_bufs);
2209 			if (cur_size < RX_LOW_JUMBO_THRES) {
2210 				if ((cur_size < RX_PANIC_JUMBO_THRES) &&
2211 				    !test_and_set_bit(0,
2212 						      &ap->jumbo_refill_busy)){
2213 #ifdef DEBUG
2214 					printk("low on jumbo buffers %i\n",
2215 					       cur_size);
2216 #endif
2217 					ace_load_jumbo_rx_ring(dev,
2218 							       RX_JUMBO_SIZE - cur_size);
2219 				} else
2220 					run_tasklet = 1;
2221 			}
2222 		}
2223 		if (run_tasklet && !ap->tasklet_pending) {
2224 			ap->tasklet_pending = 1;
2225 			tasklet_schedule(&ap->ace_tasklet);
2226 		}
2227 	}
2228 
2229 	return IRQ_HANDLED;
2230 }
2231 
2232 static int ace_open(struct net_device *dev)
2233 {
2234 	struct ace_private *ap = netdev_priv(dev);
2235 	struct ace_regs __iomem *regs = ap->regs;
2236 	struct cmd cmd;
2237 
2238 	if (!(ap->fw_running)) {
2239 		printk(KERN_WARNING "%s: Firmware not running!\n", dev->name);
2240 		return -EBUSY;
2241 	}
2242 
2243 	writel(dev->mtu + ETH_HLEN + 4, &regs->IfMtu);
2244 
2245 	cmd.evt = C_CLEAR_STATS;
2246 	cmd.code = 0;
2247 	cmd.idx = 0;
2248 	ace_issue_cmd(regs, &cmd);
2249 
2250 	cmd.evt = C_HOST_STATE;
2251 	cmd.code = C_C_STACK_UP;
2252 	cmd.idx = 0;
2253 	ace_issue_cmd(regs, &cmd);
2254 
2255 	if (ap->jumbo &&
2256 	    !test_and_set_bit(0, &ap->jumbo_refill_busy))
2257 		ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE);
2258 
2259 	if (dev->flags & IFF_PROMISC) {
2260 		cmd.evt = C_SET_PROMISC_MODE;
2261 		cmd.code = C_C_PROMISC_ENABLE;
2262 		cmd.idx = 0;
2263 		ace_issue_cmd(regs, &cmd);
2264 
2265 		ap->promisc = 1;
2266 	}else
2267 		ap->promisc = 0;
2268 	ap->mcast_all = 0;
2269 
2270 #if 0
2271 	cmd.evt = C_LNK_NEGOTIATION;
2272 	cmd.code = 0;
2273 	cmd.idx = 0;
2274 	ace_issue_cmd(regs, &cmd);
2275 #endif
2276 
2277 	netif_start_queue(dev);
2278 
2279 	/*
2280 	 * Setup the bottom half rx ring refill handler
2281 	 */
2282 	tasklet_init(&ap->ace_tasklet, ace_tasklet, (unsigned long)dev);
2283 	return 0;
2284 }
2285 
2286 
2287 static int ace_close(struct net_device *dev)
2288 {
2289 	struct ace_private *ap = netdev_priv(dev);
2290 	struct ace_regs __iomem *regs = ap->regs;
2291 	struct cmd cmd;
2292 	unsigned long flags;
2293 	short i;
2294 
2295 	/*
2296 	 * Without (or before) releasing irq and stopping hardware, this
2297 	 * is an absolute non-sense, by the way. It will be reset instantly
2298 	 * by the first irq.
2299 	 */
2300 	netif_stop_queue(dev);
2301 
2302 
2303 	if (ap->promisc) {
2304 		cmd.evt = C_SET_PROMISC_MODE;
2305 		cmd.code = C_C_PROMISC_DISABLE;
2306 		cmd.idx = 0;
2307 		ace_issue_cmd(regs, &cmd);
2308 		ap->promisc = 0;
2309 	}
2310 
2311 	cmd.evt = C_HOST_STATE;
2312 	cmd.code = C_C_STACK_DOWN;
2313 	cmd.idx = 0;
2314 	ace_issue_cmd(regs, &cmd);
2315 
2316 	tasklet_kill(&ap->ace_tasklet);
2317 
2318 	/*
2319 	 * Make sure one CPU is not processing packets while
2320 	 * buffers are being released by another.
2321 	 */
2322 
2323 	local_irq_save(flags);
2324 	ace_mask_irq(dev);
2325 
2326 	for (i = 0; i < ACE_TX_RING_ENTRIES(ap); i++) {
2327 		struct sk_buff *skb;
2328 		struct tx_ring_info *info;
2329 
2330 		info = ap->skb->tx_skbuff + i;
2331 		skb = info->skb;
2332 
2333 		if (dma_unmap_len(info, maplen)) {
2334 			if (ACE_IS_TIGON_I(ap)) {
2335 				/* NB: TIGON_1 is special, tx_ring is in io space */
2336 				struct tx_desc __iomem *tx;
2337 				tx = (__force struct tx_desc __iomem *) &ap->tx_ring[i];
2338 				writel(0, &tx->addr.addrhi);
2339 				writel(0, &tx->addr.addrlo);
2340 				writel(0, &tx->flagsize);
2341 			} else
2342 				memset(ap->tx_ring + i, 0,
2343 				       sizeof(struct tx_desc));
2344 			pci_unmap_page(ap->pdev, dma_unmap_addr(info, mapping),
2345 				       dma_unmap_len(info, maplen),
2346 				       PCI_DMA_TODEVICE);
2347 			dma_unmap_len_set(info, maplen, 0);
2348 		}
2349 		if (skb) {
2350 			dev_kfree_skb(skb);
2351 			info->skb = NULL;
2352 		}
2353 	}
2354 
2355 	if (ap->jumbo) {
2356 		cmd.evt = C_RESET_JUMBO_RNG;
2357 		cmd.code = 0;
2358 		cmd.idx = 0;
2359 		ace_issue_cmd(regs, &cmd);
2360 	}
2361 
2362 	ace_unmask_irq(dev);
2363 	local_irq_restore(flags);
2364 
2365 	return 0;
2366 }
2367 
2368 
2369 static inline dma_addr_t
2370 ace_map_tx_skb(struct ace_private *ap, struct sk_buff *skb,
2371 	       struct sk_buff *tail, u32 idx)
2372 {
2373 	dma_addr_t mapping;
2374 	struct tx_ring_info *info;
2375 
2376 	mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
2377 			       offset_in_page(skb->data),
2378 			       skb->len, PCI_DMA_TODEVICE);
2379 
2380 	info = ap->skb->tx_skbuff + idx;
2381 	info->skb = tail;
2382 	dma_unmap_addr_set(info, mapping, mapping);
2383 	dma_unmap_len_set(info, maplen, skb->len);
2384 	return mapping;
2385 }
2386 
2387 
2388 static inline void
2389 ace_load_tx_bd(struct ace_private *ap, struct tx_desc *desc, u64 addr,
2390 	       u32 flagsize, u32 vlan_tag)
2391 {
2392 #if !USE_TX_COAL_NOW
2393 	flagsize &= ~BD_FLG_COAL_NOW;
2394 #endif
2395 
2396 	if (ACE_IS_TIGON_I(ap)) {
2397 		struct tx_desc __iomem *io = (__force struct tx_desc __iomem *) desc;
2398 		writel(addr >> 32, &io->addr.addrhi);
2399 		writel(addr & 0xffffffff, &io->addr.addrlo);
2400 		writel(flagsize, &io->flagsize);
2401 		writel(vlan_tag, &io->vlanres);
2402 	} else {
2403 		desc->addr.addrhi = addr >> 32;
2404 		desc->addr.addrlo = addr;
2405 		desc->flagsize = flagsize;
2406 		desc->vlanres = vlan_tag;
2407 	}
2408 }
2409 
2410 
2411 static netdev_tx_t ace_start_xmit(struct sk_buff *skb,
2412 				  struct net_device *dev)
2413 {
2414 	struct ace_private *ap = netdev_priv(dev);
2415 	struct ace_regs __iomem *regs = ap->regs;
2416 	struct tx_desc *desc;
2417 	u32 idx, flagsize;
2418 	unsigned long maxjiff = jiffies + 3*HZ;
2419 
2420 restart:
2421 	idx = ap->tx_prd;
2422 
2423 	if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2424 		goto overflow;
2425 
2426 	if (!skb_shinfo(skb)->nr_frags)	{
2427 		dma_addr_t mapping;
2428 		u32 vlan_tag = 0;
2429 
2430 		mapping = ace_map_tx_skb(ap, skb, skb, idx);
2431 		flagsize = (skb->len << 16) | (BD_FLG_END);
2432 		if (skb->ip_summed == CHECKSUM_PARTIAL)
2433 			flagsize |= BD_FLG_TCP_UDP_SUM;
2434 		if (skb_vlan_tag_present(skb)) {
2435 			flagsize |= BD_FLG_VLAN_TAG;
2436 			vlan_tag = skb_vlan_tag_get(skb);
2437 		}
2438 		desc = ap->tx_ring + idx;
2439 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2440 
2441 		/* Look at ace_tx_int for explanations. */
2442 		if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2443 			flagsize |= BD_FLG_COAL_NOW;
2444 
2445 		ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2446 	} else {
2447 		dma_addr_t mapping;
2448 		u32 vlan_tag = 0;
2449 		int i, len = 0;
2450 
2451 		mapping = ace_map_tx_skb(ap, skb, NULL, idx);
2452 		flagsize = (skb_headlen(skb) << 16);
2453 		if (skb->ip_summed == CHECKSUM_PARTIAL)
2454 			flagsize |= BD_FLG_TCP_UDP_SUM;
2455 		if (skb_vlan_tag_present(skb)) {
2456 			flagsize |= BD_FLG_VLAN_TAG;
2457 			vlan_tag = skb_vlan_tag_get(skb);
2458 		}
2459 
2460 		ace_load_tx_bd(ap, ap->tx_ring + idx, mapping, flagsize, vlan_tag);
2461 
2462 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2463 
2464 		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2465 			const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2466 			struct tx_ring_info *info;
2467 
2468 			len += skb_frag_size(frag);
2469 			info = ap->skb->tx_skbuff + idx;
2470 			desc = ap->tx_ring + idx;
2471 
2472 			mapping = skb_frag_dma_map(&ap->pdev->dev, frag, 0,
2473 						   skb_frag_size(frag),
2474 						   DMA_TO_DEVICE);
2475 
2476 			flagsize = skb_frag_size(frag) << 16;
2477 			if (skb->ip_summed == CHECKSUM_PARTIAL)
2478 				flagsize |= BD_FLG_TCP_UDP_SUM;
2479 			idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2480 
2481 			if (i == skb_shinfo(skb)->nr_frags - 1) {
2482 				flagsize |= BD_FLG_END;
2483 				if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2484 					flagsize |= BD_FLG_COAL_NOW;
2485 
2486 				/*
2487 				 * Only the last fragment frees
2488 				 * the skb!
2489 				 */
2490 				info->skb = skb;
2491 			} else {
2492 				info->skb = NULL;
2493 			}
2494 			dma_unmap_addr_set(info, mapping, mapping);
2495 			dma_unmap_len_set(info, maplen, skb_frag_size(frag));
2496 			ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2497 		}
2498 	}
2499 
2500  	wmb();
2501  	ap->tx_prd = idx;
2502  	ace_set_txprd(regs, ap, idx);
2503 
2504 	if (flagsize & BD_FLG_COAL_NOW) {
2505 		netif_stop_queue(dev);
2506 
2507 		/*
2508 		 * A TX-descriptor producer (an IRQ) might have gotten
2509 		 * between, making the ring free again. Since xmit is
2510 		 * serialized, this is the only situation we have to
2511 		 * re-test.
2512 		 */
2513 		if (!tx_ring_full(ap, ap->tx_ret_csm, idx))
2514 			netif_wake_queue(dev);
2515 	}
2516 
2517 	return NETDEV_TX_OK;
2518 
2519 overflow:
2520 	/*
2521 	 * This race condition is unavoidable with lock-free drivers.
2522 	 * We wake up the queue _before_ tx_prd is advanced, so that we can
2523 	 * enter hard_start_xmit too early, while tx ring still looks closed.
2524 	 * This happens ~1-4 times per 100000 packets, so that we can allow
2525 	 * to loop syncing to other CPU. Probably, we need an additional
2526 	 * wmb() in ace_tx_intr as well.
2527 	 *
2528 	 * Note that this race is relieved by reserving one more entry
2529 	 * in tx ring than it is necessary (see original non-SG driver).
2530 	 * However, with SG we need to reserve 2*MAX_SKB_FRAGS+1, which
2531 	 * is already overkill.
2532 	 *
2533 	 * Alternative is to return with 1 not throttling queue. In this
2534 	 * case loop becomes longer, no more useful effects.
2535 	 */
2536 	if (time_before(jiffies, maxjiff)) {
2537 		barrier();
2538 		cpu_relax();
2539 		goto restart;
2540 	}
2541 
2542 	/* The ring is stuck full. */
2543 	printk(KERN_WARNING "%s: Transmit ring stuck full\n", dev->name);
2544 	return NETDEV_TX_BUSY;
2545 }
2546 
2547 
2548 static int ace_change_mtu(struct net_device *dev, int new_mtu)
2549 {
2550 	struct ace_private *ap = netdev_priv(dev);
2551 	struct ace_regs __iomem *regs = ap->regs;
2552 
2553 	writel(new_mtu + ETH_HLEN + 4, &regs->IfMtu);
2554 	dev->mtu = new_mtu;
2555 
2556 	if (new_mtu > ACE_STD_MTU) {
2557 		if (!(ap->jumbo)) {
2558 			printk(KERN_INFO "%s: Enabling Jumbo frame "
2559 			       "support\n", dev->name);
2560 			ap->jumbo = 1;
2561 			if (!test_and_set_bit(0, &ap->jumbo_refill_busy))
2562 				ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE);
2563 			ace_set_rxtx_parms(dev, 1);
2564 		}
2565 	} else {
2566 		while (test_and_set_bit(0, &ap->jumbo_refill_busy));
2567 		ace_sync_irq(dev->irq);
2568 		ace_set_rxtx_parms(dev, 0);
2569 		if (ap->jumbo) {
2570 			struct cmd cmd;
2571 
2572 			cmd.evt = C_RESET_JUMBO_RNG;
2573 			cmd.code = 0;
2574 			cmd.idx = 0;
2575 			ace_issue_cmd(regs, &cmd);
2576 		}
2577 	}
2578 
2579 	return 0;
2580 }
2581 
2582 static int ace_get_link_ksettings(struct net_device *dev,
2583 				  struct ethtool_link_ksettings *cmd)
2584 {
2585 	struct ace_private *ap = netdev_priv(dev);
2586 	struct ace_regs __iomem *regs = ap->regs;
2587 	u32 link;
2588 	u32 supported;
2589 
2590 	memset(cmd, 0, sizeof(struct ethtool_link_ksettings));
2591 
2592 	supported = (SUPPORTED_10baseT_Half | SUPPORTED_10baseT_Full |
2593 		     SUPPORTED_100baseT_Half | SUPPORTED_100baseT_Full |
2594 		     SUPPORTED_1000baseT_Half | SUPPORTED_1000baseT_Full |
2595 		     SUPPORTED_Autoneg | SUPPORTED_FIBRE);
2596 
2597 	cmd->base.port = PORT_FIBRE;
2598 
2599 	link = readl(&regs->GigLnkState);
2600 	if (link & LNK_1000MB) {
2601 		cmd->base.speed = SPEED_1000;
2602 	} else {
2603 		link = readl(&regs->FastLnkState);
2604 		if (link & LNK_100MB)
2605 			cmd->base.speed = SPEED_100;
2606 		else if (link & LNK_10MB)
2607 			cmd->base.speed = SPEED_10;
2608 		else
2609 			cmd->base.speed = 0;
2610 	}
2611 	if (link & LNK_FULL_DUPLEX)
2612 		cmd->base.duplex = DUPLEX_FULL;
2613 	else
2614 		cmd->base.duplex = DUPLEX_HALF;
2615 
2616 	if (link & LNK_NEGOTIATE)
2617 		cmd->base.autoneg = AUTONEG_ENABLE;
2618 	else
2619 		cmd->base.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 
2631 	ethtool_convert_legacy_u32_to_link_mode(cmd->link_modes.supported,
2632 						supported);
2633 
2634 	return 0;
2635 }
2636 
2637 static int ace_set_link_ksettings(struct net_device *dev,
2638 				  const struct ethtool_link_ksettings *cmd)
2639 {
2640 	struct ace_private *ap = netdev_priv(dev);
2641 	struct ace_regs __iomem *regs = ap->regs;
2642 	u32 link, speed;
2643 
2644 	link = readl(&regs->GigLnkState);
2645 	if (link & LNK_1000MB)
2646 		speed = SPEED_1000;
2647 	else {
2648 		link = readl(&regs->FastLnkState);
2649 		if (link & LNK_100MB)
2650 			speed = SPEED_100;
2651 		else if (link & LNK_10MB)
2652 			speed = SPEED_10;
2653 		else
2654 			speed = SPEED_100;
2655 	}
2656 
2657 	link = LNK_ENABLE | LNK_1000MB | LNK_100MB | LNK_10MB |
2658 		LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL;
2659 	if (!ACE_IS_TIGON_I(ap))
2660 		link |= LNK_TX_FLOW_CTL_Y;
2661 	if (cmd->base.autoneg == AUTONEG_ENABLE)
2662 		link |= LNK_NEGOTIATE;
2663 	if (cmd->base.speed != speed) {
2664 		link &= ~(LNK_1000MB | LNK_100MB | LNK_10MB);
2665 		switch (cmd->base.speed) {
2666 		case SPEED_1000:
2667 			link |= LNK_1000MB;
2668 			break;
2669 		case SPEED_100:
2670 			link |= LNK_100MB;
2671 			break;
2672 		case SPEED_10:
2673 			link |= LNK_10MB;
2674 			break;
2675 		}
2676 	}
2677 
2678 	if (cmd->base.duplex == DUPLEX_FULL)
2679 		link |= LNK_FULL_DUPLEX;
2680 
2681 	if (link != ap->link) {
2682 		struct cmd cmd;
2683 		printk(KERN_INFO "%s: Renegotiating link state\n",
2684 		       dev->name);
2685 
2686 		ap->link = link;
2687 		writel(link, &regs->TuneLink);
2688 		if (!ACE_IS_TIGON_I(ap))
2689 			writel(link, &regs->TuneFastLink);
2690 		wmb();
2691 
2692 		cmd.evt = C_LNK_NEGOTIATION;
2693 		cmd.code = 0;
2694 		cmd.idx = 0;
2695 		ace_issue_cmd(regs, &cmd);
2696 	}
2697 	return 0;
2698 }
2699 
2700 static void ace_get_drvinfo(struct net_device *dev,
2701 			    struct ethtool_drvinfo *info)
2702 {
2703 	struct ace_private *ap = netdev_priv(dev);
2704 
2705 	strlcpy(info->driver, "acenic", sizeof(info->driver));
2706 	snprintf(info->version, sizeof(info->version), "%i.%i.%i",
2707 		 ap->firmware_major, ap->firmware_minor,
2708 		 ap->firmware_fix);
2709 
2710 	if (ap->pdev)
2711 		strlcpy(info->bus_info, pci_name(ap->pdev),
2712 			sizeof(info->bus_info));
2713 
2714 }
2715 
2716 /*
2717  * Set the hardware MAC address.
2718  */
2719 static int ace_set_mac_addr(struct net_device *dev, void *p)
2720 {
2721 	struct ace_private *ap = netdev_priv(dev);
2722 	struct ace_regs __iomem *regs = ap->regs;
2723 	struct sockaddr *addr=p;
2724 	u8 *da;
2725 	struct cmd cmd;
2726 
2727 	if(netif_running(dev))
2728 		return -EBUSY;
2729 
2730 	memcpy(dev->dev_addr, addr->sa_data,dev->addr_len);
2731 
2732 	da = (u8 *)dev->dev_addr;
2733 
2734 	writel(da[0] << 8 | da[1], &regs->MacAddrHi);
2735 	writel((da[2] << 24) | (da[3] << 16) | (da[4] << 8) | da[5],
2736 	       &regs->MacAddrLo);
2737 
2738 	cmd.evt = C_SET_MAC_ADDR;
2739 	cmd.code = 0;
2740 	cmd.idx = 0;
2741 	ace_issue_cmd(regs, &cmd);
2742 
2743 	return 0;
2744 }
2745 
2746 
2747 static void ace_set_multicast_list(struct net_device *dev)
2748 {
2749 	struct ace_private *ap = netdev_priv(dev);
2750 	struct ace_regs __iomem *regs = ap->regs;
2751 	struct cmd cmd;
2752 
2753 	if ((dev->flags & IFF_ALLMULTI) && !(ap->mcast_all)) {
2754 		cmd.evt = C_SET_MULTICAST_MODE;
2755 		cmd.code = C_C_MCAST_ENABLE;
2756 		cmd.idx = 0;
2757 		ace_issue_cmd(regs, &cmd);
2758 		ap->mcast_all = 1;
2759 	} else if (ap->mcast_all) {
2760 		cmd.evt = C_SET_MULTICAST_MODE;
2761 		cmd.code = C_C_MCAST_DISABLE;
2762 		cmd.idx = 0;
2763 		ace_issue_cmd(regs, &cmd);
2764 		ap->mcast_all = 0;
2765 	}
2766 
2767 	if ((dev->flags & IFF_PROMISC) && !(ap->promisc)) {
2768 		cmd.evt = C_SET_PROMISC_MODE;
2769 		cmd.code = C_C_PROMISC_ENABLE;
2770 		cmd.idx = 0;
2771 		ace_issue_cmd(regs, &cmd);
2772 		ap->promisc = 1;
2773 	}else if (!(dev->flags & IFF_PROMISC) && (ap->promisc)) {
2774 		cmd.evt = C_SET_PROMISC_MODE;
2775 		cmd.code = C_C_PROMISC_DISABLE;
2776 		cmd.idx = 0;
2777 		ace_issue_cmd(regs, &cmd);
2778 		ap->promisc = 0;
2779 	}
2780 
2781 	/*
2782 	 * For the time being multicast relies on the upper layers
2783 	 * filtering it properly. The Firmware does not allow one to
2784 	 * set the entire multicast list at a time and keeping track of
2785 	 * it here is going to be messy.
2786 	 */
2787 	if (!netdev_mc_empty(dev) && !ap->mcast_all) {
2788 		cmd.evt = C_SET_MULTICAST_MODE;
2789 		cmd.code = C_C_MCAST_ENABLE;
2790 		cmd.idx = 0;
2791 		ace_issue_cmd(regs, &cmd);
2792 	}else if (!ap->mcast_all) {
2793 		cmd.evt = C_SET_MULTICAST_MODE;
2794 		cmd.code = C_C_MCAST_DISABLE;
2795 		cmd.idx = 0;
2796 		ace_issue_cmd(regs, &cmd);
2797 	}
2798 }
2799 
2800 
2801 static struct net_device_stats *ace_get_stats(struct net_device *dev)
2802 {
2803 	struct ace_private *ap = netdev_priv(dev);
2804 	struct ace_mac_stats __iomem *mac_stats =
2805 		(struct ace_mac_stats __iomem *)ap->regs->Stats;
2806 
2807 	dev->stats.rx_missed_errors = readl(&mac_stats->drop_space);
2808 	dev->stats.multicast = readl(&mac_stats->kept_mc);
2809 	dev->stats.collisions = readl(&mac_stats->coll);
2810 
2811 	return &dev->stats;
2812 }
2813 
2814 
2815 static void ace_copy(struct ace_regs __iomem *regs, const __be32 *src,
2816 		     u32 dest, int size)
2817 {
2818 	void __iomem *tdest;
2819 	short tsize, i;
2820 
2821 	if (size <= 0)
2822 		return;
2823 
2824 	while (size > 0) {
2825 		tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2826 			    min_t(u32, size, ACE_WINDOW_SIZE));
2827 		tdest = (void __iomem *) &regs->Window +
2828 			(dest & (ACE_WINDOW_SIZE - 1));
2829 		writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2830 		for (i = 0; i < (tsize / 4); i++) {
2831 			/* Firmware is big-endian */
2832 			writel(be32_to_cpup(src), tdest);
2833 			src++;
2834 			tdest += 4;
2835 			dest += 4;
2836 			size -= 4;
2837 		}
2838 	}
2839 }
2840 
2841 
2842 static void ace_clear(struct ace_regs __iomem *regs, u32 dest, int size)
2843 {
2844 	void __iomem *tdest;
2845 	short tsize = 0, i;
2846 
2847 	if (size <= 0)
2848 		return;
2849 
2850 	while (size > 0) {
2851 		tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2852 				min_t(u32, size, ACE_WINDOW_SIZE));
2853 		tdest = (void __iomem *) &regs->Window +
2854 			(dest & (ACE_WINDOW_SIZE - 1));
2855 		writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2856 
2857 		for (i = 0; i < (tsize / 4); i++) {
2858 			writel(0, tdest + i*4);
2859 		}
2860 
2861 		dest += tsize;
2862 		size -= tsize;
2863 	}
2864 }
2865 
2866 
2867 /*
2868  * Download the firmware into the SRAM on the NIC
2869  *
2870  * This operation requires the NIC to be halted and is performed with
2871  * interrupts disabled and with the spinlock hold.
2872  */
2873 static int ace_load_firmware(struct net_device *dev)
2874 {
2875 	const struct firmware *fw;
2876 	const char *fw_name = "acenic/tg2.bin";
2877 	struct ace_private *ap = netdev_priv(dev);
2878 	struct ace_regs __iomem *regs = ap->regs;
2879 	const __be32 *fw_data;
2880 	u32 load_addr;
2881 	int ret;
2882 
2883 	if (!(readl(&regs->CpuCtrl) & CPU_HALTED)) {
2884 		printk(KERN_ERR "%s: trying to download firmware while the "
2885 		       "CPU is running!\n", ap->name);
2886 		return -EFAULT;
2887 	}
2888 
2889 	if (ACE_IS_TIGON_I(ap))
2890 		fw_name = "acenic/tg1.bin";
2891 
2892 	ret = request_firmware(&fw, fw_name, &ap->pdev->dev);
2893 	if (ret) {
2894 		printk(KERN_ERR "%s: Failed to load firmware \"%s\"\n",
2895 		       ap->name, fw_name);
2896 		return ret;
2897 	}
2898 
2899 	fw_data = (void *)fw->data;
2900 
2901 	/* Firmware blob starts with version numbers, followed by
2902 	   load and start address. Remainder is the blob to be loaded
2903 	   contiguously from load address. We don't bother to represent
2904 	   the BSS/SBSS sections any more, since we were clearing the
2905 	   whole thing anyway. */
2906 	ap->firmware_major = fw->data[0];
2907 	ap->firmware_minor = fw->data[1];
2908 	ap->firmware_fix = fw->data[2];
2909 
2910 	ap->firmware_start = be32_to_cpu(fw_data[1]);
2911 	if (ap->firmware_start < 0x4000 || ap->firmware_start >= 0x80000) {
2912 		printk(KERN_ERR "%s: bogus load address %08x in \"%s\"\n",
2913 		       ap->name, ap->firmware_start, fw_name);
2914 		ret = -EINVAL;
2915 		goto out;
2916 	}
2917 
2918 	load_addr = be32_to_cpu(fw_data[2]);
2919 	if (load_addr < 0x4000 || load_addr >= 0x80000) {
2920 		printk(KERN_ERR "%s: bogus load address %08x in \"%s\"\n",
2921 		       ap->name, load_addr, fw_name);
2922 		ret = -EINVAL;
2923 		goto out;
2924 	}
2925 
2926 	/*
2927 	 * Do not try to clear more than 512KiB or we end up seeing
2928 	 * funny things on NICs with only 512KiB SRAM
2929 	 */
2930 	ace_clear(regs, 0x2000, 0x80000-0x2000);
2931 	ace_copy(regs, &fw_data[3], load_addr, fw->size-12);
2932  out:
2933 	release_firmware(fw);
2934 	return ret;
2935 }
2936 
2937 
2938 /*
2939  * The eeprom on the AceNIC is an Atmel i2c EEPROM.
2940  *
2941  * Accessing the EEPROM is `interesting' to say the least - don't read
2942  * this code right after dinner.
2943  *
2944  * This is all about black magic and bit-banging the device .... I
2945  * wonder in what hospital they have put the guy who designed the i2c
2946  * specs.
2947  *
2948  * Oh yes, this is only the beginning!
2949  *
2950  * Thanks to Stevarino Webinski for helping tracking down the bugs in the
2951  * code i2c readout code by beta testing all my hacks.
2952  */
2953 static void eeprom_start(struct ace_regs __iomem *regs)
2954 {
2955 	u32 local;
2956 
2957 	readl(&regs->LocalCtrl);
2958 	udelay(ACE_SHORT_DELAY);
2959 	local = readl(&regs->LocalCtrl);
2960 	local |= EEPROM_DATA_OUT | EEPROM_WRITE_ENABLE;
2961 	writel(local, &regs->LocalCtrl);
2962 	readl(&regs->LocalCtrl);
2963 	mb();
2964 	udelay(ACE_SHORT_DELAY);
2965 	local |= EEPROM_CLK_OUT;
2966 	writel(local, &regs->LocalCtrl);
2967 	readl(&regs->LocalCtrl);
2968 	mb();
2969 	udelay(ACE_SHORT_DELAY);
2970 	local &= ~EEPROM_DATA_OUT;
2971 	writel(local, &regs->LocalCtrl);
2972 	readl(&regs->LocalCtrl);
2973 	mb();
2974 	udelay(ACE_SHORT_DELAY);
2975 	local &= ~EEPROM_CLK_OUT;
2976 	writel(local, &regs->LocalCtrl);
2977 	readl(&regs->LocalCtrl);
2978 	mb();
2979 }
2980 
2981 
2982 static void eeprom_prep(struct ace_regs __iomem *regs, u8 magic)
2983 {
2984 	short i;
2985 	u32 local;
2986 
2987 	udelay(ACE_SHORT_DELAY);
2988 	local = readl(&regs->LocalCtrl);
2989 	local &= ~EEPROM_DATA_OUT;
2990 	local |= EEPROM_WRITE_ENABLE;
2991 	writel(local, &regs->LocalCtrl);
2992 	readl(&regs->LocalCtrl);
2993 	mb();
2994 
2995 	for (i = 0; i < 8; i++, magic <<= 1) {
2996 		udelay(ACE_SHORT_DELAY);
2997 		if (magic & 0x80)
2998 			local |= EEPROM_DATA_OUT;
2999 		else
3000 			local &= ~EEPROM_DATA_OUT;
3001 		writel(local, &regs->LocalCtrl);
3002 		readl(&regs->LocalCtrl);
3003 		mb();
3004 
3005 		udelay(ACE_SHORT_DELAY);
3006 		local |= EEPROM_CLK_OUT;
3007 		writel(local, &regs->LocalCtrl);
3008 		readl(&regs->LocalCtrl);
3009 		mb();
3010 		udelay(ACE_SHORT_DELAY);
3011 		local &= ~(EEPROM_CLK_OUT | EEPROM_DATA_OUT);
3012 		writel(local, &regs->LocalCtrl);
3013 		readl(&regs->LocalCtrl);
3014 		mb();
3015 	}
3016 }
3017 
3018 
3019 static int eeprom_check_ack(struct ace_regs __iomem *regs)
3020 {
3021 	int state;
3022 	u32 local;
3023 
3024 	local = readl(&regs->LocalCtrl);
3025 	local &= ~EEPROM_WRITE_ENABLE;
3026 	writel(local, &regs->LocalCtrl);
3027 	readl(&regs->LocalCtrl);
3028 	mb();
3029 	udelay(ACE_LONG_DELAY);
3030 	local |= EEPROM_CLK_OUT;
3031 	writel(local, &regs->LocalCtrl);
3032 	readl(&regs->LocalCtrl);
3033 	mb();
3034 	udelay(ACE_SHORT_DELAY);
3035 	/* sample data in middle of high clk */
3036 	state = (readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0;
3037 	udelay(ACE_SHORT_DELAY);
3038 	mb();
3039 	writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3040 	readl(&regs->LocalCtrl);
3041 	mb();
3042 
3043 	return state;
3044 }
3045 
3046 
3047 static void eeprom_stop(struct ace_regs __iomem *regs)
3048 {
3049 	u32 local;
3050 
3051 	udelay(ACE_SHORT_DELAY);
3052 	local = readl(&regs->LocalCtrl);
3053 	local |= EEPROM_WRITE_ENABLE;
3054 	writel(local, &regs->LocalCtrl);
3055 	readl(&regs->LocalCtrl);
3056 	mb();
3057 	udelay(ACE_SHORT_DELAY);
3058 	local &= ~EEPROM_DATA_OUT;
3059 	writel(local, &regs->LocalCtrl);
3060 	readl(&regs->LocalCtrl);
3061 	mb();
3062 	udelay(ACE_SHORT_DELAY);
3063 	local |= EEPROM_CLK_OUT;
3064 	writel(local, &regs->LocalCtrl);
3065 	readl(&regs->LocalCtrl);
3066 	mb();
3067 	udelay(ACE_SHORT_DELAY);
3068 	local |= EEPROM_DATA_OUT;
3069 	writel(local, &regs->LocalCtrl);
3070 	readl(&regs->LocalCtrl);
3071 	mb();
3072 	udelay(ACE_LONG_DELAY);
3073 	local &= ~EEPROM_CLK_OUT;
3074 	writel(local, &regs->LocalCtrl);
3075 	mb();
3076 }
3077 
3078 
3079 /*
3080  * Read a whole byte from the EEPROM.
3081  */
3082 static int read_eeprom_byte(struct net_device *dev, unsigned long offset)
3083 {
3084 	struct ace_private *ap = netdev_priv(dev);
3085 	struct ace_regs __iomem *regs = ap->regs;
3086 	unsigned long flags;
3087 	u32 local;
3088 	int result = 0;
3089 	short i;
3090 
3091 	/*
3092 	 * Don't take interrupts on this CPU will bit banging
3093 	 * the %#%#@$ I2C device
3094 	 */
3095 	local_irq_save(flags);
3096 
3097 	eeprom_start(regs);
3098 
3099 	eeprom_prep(regs, EEPROM_WRITE_SELECT);
3100 	if (eeprom_check_ack(regs)) {
3101 		local_irq_restore(flags);
3102 		printk(KERN_ERR "%s: Unable to sync eeprom\n", ap->name);
3103 		result = -EIO;
3104 		goto eeprom_read_error;
3105 	}
3106 
3107 	eeprom_prep(regs, (offset >> 8) & 0xff);
3108 	if (eeprom_check_ack(regs)) {
3109 		local_irq_restore(flags);
3110 		printk(KERN_ERR "%s: Unable to set address byte 0\n",
3111 		       ap->name);
3112 		result = -EIO;
3113 		goto eeprom_read_error;
3114 	}
3115 
3116 	eeprom_prep(regs, offset & 0xff);
3117 	if (eeprom_check_ack(regs)) {
3118 		local_irq_restore(flags);
3119 		printk(KERN_ERR "%s: Unable to set address byte 1\n",
3120 		       ap->name);
3121 		result = -EIO;
3122 		goto eeprom_read_error;
3123 	}
3124 
3125 	eeprom_start(regs);
3126 	eeprom_prep(regs, EEPROM_READ_SELECT);
3127 	if (eeprom_check_ack(regs)) {
3128 		local_irq_restore(flags);
3129 		printk(KERN_ERR "%s: Unable to set READ_SELECT\n",
3130 		       ap->name);
3131 		result = -EIO;
3132 		goto eeprom_read_error;
3133 	}
3134 
3135 	for (i = 0; i < 8; i++) {
3136 		local = readl(&regs->LocalCtrl);
3137 		local &= ~EEPROM_WRITE_ENABLE;
3138 		writel(local, &regs->LocalCtrl);
3139 		readl(&regs->LocalCtrl);
3140 		udelay(ACE_LONG_DELAY);
3141 		mb();
3142 		local |= EEPROM_CLK_OUT;
3143 		writel(local, &regs->LocalCtrl);
3144 		readl(&regs->LocalCtrl);
3145 		mb();
3146 		udelay(ACE_SHORT_DELAY);
3147 		/* sample data mid high clk */
3148 		result = (result << 1) |
3149 			((readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0);
3150 		udelay(ACE_SHORT_DELAY);
3151 		mb();
3152 		local = readl(&regs->LocalCtrl);
3153 		local &= ~EEPROM_CLK_OUT;
3154 		writel(local, &regs->LocalCtrl);
3155 		readl(&regs->LocalCtrl);
3156 		udelay(ACE_SHORT_DELAY);
3157 		mb();
3158 		if (i == 7) {
3159 			local |= EEPROM_WRITE_ENABLE;
3160 			writel(local, &regs->LocalCtrl);
3161 			readl(&regs->LocalCtrl);
3162 			mb();
3163 			udelay(ACE_SHORT_DELAY);
3164 		}
3165 	}
3166 
3167 	local |= EEPROM_DATA_OUT;
3168 	writel(local, &regs->LocalCtrl);
3169 	readl(&regs->LocalCtrl);
3170 	mb();
3171 	udelay(ACE_SHORT_DELAY);
3172 	writel(readl(&regs->LocalCtrl) | EEPROM_CLK_OUT, &regs->LocalCtrl);
3173 	readl(&regs->LocalCtrl);
3174 	udelay(ACE_LONG_DELAY);
3175 	writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3176 	readl(&regs->LocalCtrl);
3177 	mb();
3178 	udelay(ACE_SHORT_DELAY);
3179 	eeprom_stop(regs);
3180 
3181 	local_irq_restore(flags);
3182  out:
3183 	return result;
3184 
3185  eeprom_read_error:
3186 	printk(KERN_ERR "%s: Unable to read eeprom byte 0x%02lx\n",
3187 	       ap->name, offset);
3188 	goto out;
3189 }
3190 
3191 module_pci_driver(acenic_pci_driver);
3192