xref: /openbmc/linux/drivers/net/ethernet/alteon/acenic.c (revision a0ae2562c6c4b2721d9fddba63b7286c13517d9f)
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 	case PCI_VENDOR_ID_SGI:
555 		printk(KERN_INFO "%s: SGI AceNIC ", ap->name);
556 		break;
557 	default:
558 		printk(KERN_INFO "%s: Unknown AceNIC ", ap->name);
559 		break;
560 	}
561 
562 	printk("Gigabit Ethernet at 0x%08lx, ", dev->base_addr);
563 	printk("irq %d\n", pdev->irq);
564 
565 #ifdef CONFIG_ACENIC_OMIT_TIGON_I
566 	if ((readl(&ap->regs->HostCtrl) >> 28) == 4) {
567 		printk(KERN_ERR "%s: Driver compiled without Tigon I"
568 		       " support - NIC disabled\n", dev->name);
569 		goto fail_uninit;
570 	}
571 #endif
572 
573 	if (ace_allocate_descriptors(dev))
574 		goto fail_free_netdev;
575 
576 #ifdef MODULE
577 	if (boards_found >= ACE_MAX_MOD_PARMS)
578 		ap->board_idx = BOARD_IDX_OVERFLOW;
579 	else
580 		ap->board_idx = boards_found;
581 #else
582 	ap->board_idx = BOARD_IDX_STATIC;
583 #endif
584 
585 	if (ace_init(dev))
586 		goto fail_free_netdev;
587 
588 	if (register_netdev(dev)) {
589 		printk(KERN_ERR "acenic: device registration failed\n");
590 		goto fail_uninit;
591 	}
592 	ap->name = dev->name;
593 
594 	if (ap->pci_using_dac)
595 		dev->features |= NETIF_F_HIGHDMA;
596 
597 	pci_set_drvdata(pdev, dev);
598 
599 	boards_found++;
600 	return 0;
601 
602  fail_uninit:
603 	ace_init_cleanup(dev);
604  fail_free_netdev:
605 	free_netdev(dev);
606 	return -ENODEV;
607 }
608 
609 static void acenic_remove_one(struct pci_dev *pdev)
610 {
611 	struct net_device *dev = pci_get_drvdata(pdev);
612 	struct ace_private *ap = netdev_priv(dev);
613 	struct ace_regs __iomem *regs = ap->regs;
614 	short i;
615 
616 	unregister_netdev(dev);
617 
618 	writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
619 	if (ap->version >= 2)
620 		writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
621 
622 	/*
623 	 * This clears any pending interrupts
624 	 */
625 	writel(1, &regs->Mb0Lo);
626 	readl(&regs->CpuCtrl);	/* flush */
627 
628 	/*
629 	 * Make sure no other CPUs are processing interrupts
630 	 * on the card before the buffers are being released.
631 	 * Otherwise one might experience some `interesting'
632 	 * effects.
633 	 *
634 	 * Then release the RX buffers - jumbo buffers were
635 	 * already released in ace_close().
636 	 */
637 	ace_sync_irq(dev->irq);
638 
639 	for (i = 0; i < RX_STD_RING_ENTRIES; i++) {
640 		struct sk_buff *skb = ap->skb->rx_std_skbuff[i].skb;
641 
642 		if (skb) {
643 			struct ring_info *ringp;
644 			dma_addr_t mapping;
645 
646 			ringp = &ap->skb->rx_std_skbuff[i];
647 			mapping = dma_unmap_addr(ringp, mapping);
648 			pci_unmap_page(ap->pdev, mapping,
649 				       ACE_STD_BUFSIZE,
650 				       PCI_DMA_FROMDEVICE);
651 
652 			ap->rx_std_ring[i].size = 0;
653 			ap->skb->rx_std_skbuff[i].skb = NULL;
654 			dev_kfree_skb(skb);
655 		}
656 	}
657 
658 	if (ap->version >= 2) {
659 		for (i = 0; i < RX_MINI_RING_ENTRIES; i++) {
660 			struct sk_buff *skb = ap->skb->rx_mini_skbuff[i].skb;
661 
662 			if (skb) {
663 				struct ring_info *ringp;
664 				dma_addr_t mapping;
665 
666 				ringp = &ap->skb->rx_mini_skbuff[i];
667 				mapping = dma_unmap_addr(ringp,mapping);
668 				pci_unmap_page(ap->pdev, mapping,
669 					       ACE_MINI_BUFSIZE,
670 					       PCI_DMA_FROMDEVICE);
671 
672 				ap->rx_mini_ring[i].size = 0;
673 				ap->skb->rx_mini_skbuff[i].skb = NULL;
674 				dev_kfree_skb(skb);
675 			}
676 		}
677 	}
678 
679 	for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
680 		struct sk_buff *skb = ap->skb->rx_jumbo_skbuff[i].skb;
681 		if (skb) {
682 			struct ring_info *ringp;
683 			dma_addr_t mapping;
684 
685 			ringp = &ap->skb->rx_jumbo_skbuff[i];
686 			mapping = dma_unmap_addr(ringp, mapping);
687 			pci_unmap_page(ap->pdev, mapping,
688 				       ACE_JUMBO_BUFSIZE,
689 				       PCI_DMA_FROMDEVICE);
690 
691 			ap->rx_jumbo_ring[i].size = 0;
692 			ap->skb->rx_jumbo_skbuff[i].skb = NULL;
693 			dev_kfree_skb(skb);
694 		}
695 	}
696 
697 	ace_init_cleanup(dev);
698 	free_netdev(dev);
699 }
700 
701 static struct pci_driver acenic_pci_driver = {
702 	.name		= "acenic",
703 	.id_table	= acenic_pci_tbl,
704 	.probe		= acenic_probe_one,
705 	.remove		= acenic_remove_one,
706 };
707 
708 static void ace_free_descriptors(struct net_device *dev)
709 {
710 	struct ace_private *ap = netdev_priv(dev);
711 	int size;
712 
713 	if (ap->rx_std_ring != NULL) {
714 		size = (sizeof(struct rx_desc) *
715 			(RX_STD_RING_ENTRIES +
716 			 RX_JUMBO_RING_ENTRIES +
717 			 RX_MINI_RING_ENTRIES +
718 			 RX_RETURN_RING_ENTRIES));
719 		pci_free_consistent(ap->pdev, size, ap->rx_std_ring,
720 				    ap->rx_ring_base_dma);
721 		ap->rx_std_ring = NULL;
722 		ap->rx_jumbo_ring = NULL;
723 		ap->rx_mini_ring = NULL;
724 		ap->rx_return_ring = NULL;
725 	}
726 	if (ap->evt_ring != NULL) {
727 		size = (sizeof(struct event) * EVT_RING_ENTRIES);
728 		pci_free_consistent(ap->pdev, size, ap->evt_ring,
729 				    ap->evt_ring_dma);
730 		ap->evt_ring = NULL;
731 	}
732 	if (ap->tx_ring != NULL && !ACE_IS_TIGON_I(ap)) {
733 		size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
734 		pci_free_consistent(ap->pdev, size, ap->tx_ring,
735 				    ap->tx_ring_dma);
736 	}
737 	ap->tx_ring = NULL;
738 
739 	if (ap->evt_prd != NULL) {
740 		pci_free_consistent(ap->pdev, sizeof(u32),
741 				    (void *)ap->evt_prd, ap->evt_prd_dma);
742 		ap->evt_prd = NULL;
743 	}
744 	if (ap->rx_ret_prd != NULL) {
745 		pci_free_consistent(ap->pdev, sizeof(u32),
746 				    (void *)ap->rx_ret_prd,
747 				    ap->rx_ret_prd_dma);
748 		ap->rx_ret_prd = NULL;
749 	}
750 	if (ap->tx_csm != NULL) {
751 		pci_free_consistent(ap->pdev, sizeof(u32),
752 				    (void *)ap->tx_csm, ap->tx_csm_dma);
753 		ap->tx_csm = NULL;
754 	}
755 }
756 
757 
758 static int ace_allocate_descriptors(struct net_device *dev)
759 {
760 	struct ace_private *ap = netdev_priv(dev);
761 	int size;
762 
763 	size = (sizeof(struct rx_desc) *
764 		(RX_STD_RING_ENTRIES +
765 		 RX_JUMBO_RING_ENTRIES +
766 		 RX_MINI_RING_ENTRIES +
767 		 RX_RETURN_RING_ENTRIES));
768 
769 	ap->rx_std_ring = pci_alloc_consistent(ap->pdev, size,
770 					       &ap->rx_ring_base_dma);
771 	if (ap->rx_std_ring == NULL)
772 		goto fail;
773 
774 	ap->rx_jumbo_ring = ap->rx_std_ring + RX_STD_RING_ENTRIES;
775 	ap->rx_mini_ring = ap->rx_jumbo_ring + RX_JUMBO_RING_ENTRIES;
776 	ap->rx_return_ring = ap->rx_mini_ring + RX_MINI_RING_ENTRIES;
777 
778 	size = (sizeof(struct event) * EVT_RING_ENTRIES);
779 
780 	ap->evt_ring = pci_alloc_consistent(ap->pdev, size, &ap->evt_ring_dma);
781 
782 	if (ap->evt_ring == NULL)
783 		goto fail;
784 
785 	/*
786 	 * Only allocate a host TX ring for the Tigon II, the Tigon I
787 	 * has to use PCI registers for this ;-(
788 	 */
789 	if (!ACE_IS_TIGON_I(ap)) {
790 		size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
791 
792 		ap->tx_ring = pci_alloc_consistent(ap->pdev, size,
793 						   &ap->tx_ring_dma);
794 
795 		if (ap->tx_ring == NULL)
796 			goto fail;
797 	}
798 
799 	ap->evt_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
800 					   &ap->evt_prd_dma);
801 	if (ap->evt_prd == NULL)
802 		goto fail;
803 
804 	ap->rx_ret_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
805 					      &ap->rx_ret_prd_dma);
806 	if (ap->rx_ret_prd == NULL)
807 		goto fail;
808 
809 	ap->tx_csm = pci_alloc_consistent(ap->pdev, sizeof(u32),
810 					  &ap->tx_csm_dma);
811 	if (ap->tx_csm == NULL)
812 		goto fail;
813 
814 	return 0;
815 
816 fail:
817 	/* Clean up. */
818 	ace_init_cleanup(dev);
819 	return 1;
820 }
821 
822 
823 /*
824  * Generic cleanup handling data allocated during init. Used when the
825  * module is unloaded or if an error occurs during initialization
826  */
827 static void ace_init_cleanup(struct net_device *dev)
828 {
829 	struct ace_private *ap;
830 
831 	ap = netdev_priv(dev);
832 
833 	ace_free_descriptors(dev);
834 
835 	if (ap->info)
836 		pci_free_consistent(ap->pdev, sizeof(struct ace_info),
837 				    ap->info, ap->info_dma);
838 	kfree(ap->skb);
839 	kfree(ap->trace_buf);
840 
841 	if (dev->irq)
842 		free_irq(dev->irq, dev);
843 
844 	iounmap(ap->regs);
845 }
846 
847 
848 /*
849  * Commands are considered to be slow.
850  */
851 static inline void ace_issue_cmd(struct ace_regs __iomem *regs, struct cmd *cmd)
852 {
853 	u32 idx;
854 
855 	idx = readl(&regs->CmdPrd);
856 
857 	writel(*(u32 *)(cmd), &regs->CmdRng[idx]);
858 	idx = (idx + 1) % CMD_RING_ENTRIES;
859 
860 	writel(idx, &regs->CmdPrd);
861 }
862 
863 
864 static int ace_init(struct net_device *dev)
865 {
866 	struct ace_private *ap;
867 	struct ace_regs __iomem *regs;
868 	struct ace_info *info = NULL;
869 	struct pci_dev *pdev;
870 	unsigned long myjif;
871 	u64 tmp_ptr;
872 	u32 tig_ver, mac1, mac2, tmp, pci_state;
873 	int board_idx, ecode = 0;
874 	short i;
875 	unsigned char cache_size;
876 
877 	ap = netdev_priv(dev);
878 	regs = ap->regs;
879 
880 	board_idx = ap->board_idx;
881 
882 	/*
883 	 * aman@sgi.com - its useful to do a NIC reset here to
884 	 * address the `Firmware not running' problem subsequent
885 	 * to any crashes involving the NIC
886 	 */
887 	writel(HW_RESET | (HW_RESET << 24), &regs->HostCtrl);
888 	readl(&regs->HostCtrl);		/* PCI write posting */
889 	udelay(5);
890 
891 	/*
892 	 * Don't access any other registers before this point!
893 	 */
894 #ifdef __BIG_ENDIAN
895 	/*
896 	 * This will most likely need BYTE_SWAP once we switch
897 	 * to using __raw_writel()
898 	 */
899 	writel((WORD_SWAP | CLR_INT | ((WORD_SWAP | CLR_INT) << 24)),
900 	       &regs->HostCtrl);
901 #else
902 	writel((CLR_INT | WORD_SWAP | ((CLR_INT | WORD_SWAP) << 24)),
903 	       &regs->HostCtrl);
904 #endif
905 	readl(&regs->HostCtrl);		/* PCI write posting */
906 
907 	/*
908 	 * Stop the NIC CPU and clear pending interrupts
909 	 */
910 	writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
911 	readl(&regs->CpuCtrl);		/* PCI write posting */
912 	writel(0, &regs->Mb0Lo);
913 
914 	tig_ver = readl(&regs->HostCtrl) >> 28;
915 
916 	switch(tig_ver){
917 #ifndef CONFIG_ACENIC_OMIT_TIGON_I
918 	case 4:
919 	case 5:
920 		printk(KERN_INFO "  Tigon I  (Rev. %i), Firmware: %i.%i.%i, ",
921 		       tig_ver, ap->firmware_major, ap->firmware_minor,
922 		       ap->firmware_fix);
923 		writel(0, &regs->LocalCtrl);
924 		ap->version = 1;
925 		ap->tx_ring_entries = TIGON_I_TX_RING_ENTRIES;
926 		break;
927 #endif
928 	case 6:
929 		printk(KERN_INFO "  Tigon II (Rev. %i), Firmware: %i.%i.%i, ",
930 		       tig_ver, ap->firmware_major, ap->firmware_minor,
931 		       ap->firmware_fix);
932 		writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
933 		readl(&regs->CpuBCtrl);		/* PCI write posting */
934 		/*
935 		 * The SRAM bank size does _not_ indicate the amount
936 		 * of memory on the card, it controls the _bank_ size!
937 		 * Ie. a 1MB AceNIC will have two banks of 512KB.
938 		 */
939 		writel(SRAM_BANK_512K, &regs->LocalCtrl);
940 		writel(SYNC_SRAM_TIMING, &regs->MiscCfg);
941 		ap->version = 2;
942 		ap->tx_ring_entries = MAX_TX_RING_ENTRIES;
943 		break;
944 	default:
945 		printk(KERN_WARNING "  Unsupported Tigon version detected "
946 		       "(%i)\n", tig_ver);
947 		ecode = -ENODEV;
948 		goto init_error;
949 	}
950 
951 	/*
952 	 * ModeStat _must_ be set after the SRAM settings as this change
953 	 * seems to corrupt the ModeStat and possible other registers.
954 	 * The SRAM settings survive resets and setting it to the same
955 	 * value a second time works as well. This is what caused the
956 	 * `Firmware not running' problem on the Tigon II.
957 	 */
958 #ifdef __BIG_ENDIAN
959 	writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL | ACE_BYTE_SWAP_BD |
960 	       ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
961 #else
962 	writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL |
963 	       ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
964 #endif
965 	readl(&regs->ModeStat);		/* PCI write posting */
966 
967 	mac1 = 0;
968 	for(i = 0; i < 4; i++) {
969 		int t;
970 
971 		mac1 = mac1 << 8;
972 		t = read_eeprom_byte(dev, 0x8c+i);
973 		if (t < 0) {
974 			ecode = -EIO;
975 			goto init_error;
976 		} else
977 			mac1 |= (t & 0xff);
978 	}
979 	mac2 = 0;
980 	for(i = 4; i < 8; i++) {
981 		int t;
982 
983 		mac2 = mac2 << 8;
984 		t = read_eeprom_byte(dev, 0x8c+i);
985 		if (t < 0) {
986 			ecode = -EIO;
987 			goto init_error;
988 		} else
989 			mac2 |= (t & 0xff);
990 	}
991 
992 	writel(mac1, &regs->MacAddrHi);
993 	writel(mac2, &regs->MacAddrLo);
994 
995 	dev->dev_addr[0] = (mac1 >> 8) & 0xff;
996 	dev->dev_addr[1] = mac1 & 0xff;
997 	dev->dev_addr[2] = (mac2 >> 24) & 0xff;
998 	dev->dev_addr[3] = (mac2 >> 16) & 0xff;
999 	dev->dev_addr[4] = (mac2 >> 8) & 0xff;
1000 	dev->dev_addr[5] = mac2 & 0xff;
1001 
1002 	printk("MAC: %pM\n", dev->dev_addr);
1003 
1004 	/*
1005 	 * Looks like this is necessary to deal with on all architectures,
1006 	 * even this %$#%$# N440BX Intel based thing doesn't get it right.
1007 	 * Ie. having two NICs in the machine, one will have the cache
1008 	 * line set at boot time, the other will not.
1009 	 */
1010 	pdev = ap->pdev;
1011 	pci_read_config_byte(pdev, PCI_CACHE_LINE_SIZE, &cache_size);
1012 	cache_size <<= 2;
1013 	if (cache_size != SMP_CACHE_BYTES) {
1014 		printk(KERN_INFO "  PCI cache line size set incorrectly "
1015 		       "(%i bytes) by BIOS/FW, ", cache_size);
1016 		if (cache_size > SMP_CACHE_BYTES)
1017 			printk("expecting %i\n", SMP_CACHE_BYTES);
1018 		else {
1019 			printk("correcting to %i\n", SMP_CACHE_BYTES);
1020 			pci_write_config_byte(pdev, PCI_CACHE_LINE_SIZE,
1021 					      SMP_CACHE_BYTES >> 2);
1022 		}
1023 	}
1024 
1025 	pci_state = readl(&regs->PciState);
1026 	printk(KERN_INFO "  PCI bus width: %i bits, speed: %iMHz, "
1027 	       "latency: %i clks\n",
1028 	       	(pci_state & PCI_32BIT) ? 32 : 64,
1029 		(pci_state & PCI_66MHZ) ? 66 : 33,
1030 		ap->pci_latency);
1031 
1032 	/*
1033 	 * Set the max DMA transfer size. Seems that for most systems
1034 	 * the performance is better when no MAX parameter is
1035 	 * set. However for systems enabling PCI write and invalidate,
1036 	 * DMA writes must be set to the L1 cache line size to get
1037 	 * optimal performance.
1038 	 *
1039 	 * The default is now to turn the PCI write and invalidate off
1040 	 * - that is what Alteon does for NT.
1041 	 */
1042 	tmp = READ_CMD_MEM | WRITE_CMD_MEM;
1043 	if (ap->version >= 2) {
1044 		tmp |= (MEM_READ_MULTIPLE | (pci_state & PCI_66MHZ));
1045 		/*
1046 		 * Tuning parameters only supported for 8 cards
1047 		 */
1048 		if (board_idx == BOARD_IDX_OVERFLOW ||
1049 		    dis_pci_mem_inval[board_idx]) {
1050 			if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1051 				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1052 				pci_write_config_word(pdev, PCI_COMMAND,
1053 						      ap->pci_command);
1054 				printk(KERN_INFO "  Disabling PCI memory "
1055 				       "write and invalidate\n");
1056 			}
1057 		} else if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1058 			printk(KERN_INFO "  PCI memory write & invalidate "
1059 			       "enabled by BIOS, enabling counter measures\n");
1060 
1061 			switch(SMP_CACHE_BYTES) {
1062 			case 16:
1063 				tmp |= DMA_WRITE_MAX_16;
1064 				break;
1065 			case 32:
1066 				tmp |= DMA_WRITE_MAX_32;
1067 				break;
1068 			case 64:
1069 				tmp |= DMA_WRITE_MAX_64;
1070 				break;
1071 			case 128:
1072 				tmp |= DMA_WRITE_MAX_128;
1073 				break;
1074 			default:
1075 				printk(KERN_INFO "  Cache line size %i not "
1076 				       "supported, PCI write and invalidate "
1077 				       "disabled\n", SMP_CACHE_BYTES);
1078 				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1079 				pci_write_config_word(pdev, PCI_COMMAND,
1080 						      ap->pci_command);
1081 			}
1082 		}
1083 	}
1084 
1085 #ifdef __sparc__
1086 	/*
1087 	 * On this platform, we know what the best dma settings
1088 	 * are.  We use 64-byte maximum bursts, because if we
1089 	 * burst larger than the cache line size (or even cross
1090 	 * a 64byte boundary in a single burst) the UltraSparc
1091 	 * PCI controller will disconnect at 64-byte multiples.
1092 	 *
1093 	 * Read-multiple will be properly enabled above, and when
1094 	 * set will give the PCI controller proper hints about
1095 	 * prefetching.
1096 	 */
1097 	tmp &= ~DMA_READ_WRITE_MASK;
1098 	tmp |= DMA_READ_MAX_64;
1099 	tmp |= DMA_WRITE_MAX_64;
1100 #endif
1101 #ifdef __alpha__
1102 	tmp &= ~DMA_READ_WRITE_MASK;
1103 	tmp |= DMA_READ_MAX_128;
1104 	/*
1105 	 * All the docs say MUST NOT. Well, I did.
1106 	 * Nothing terrible happens, if we load wrong size.
1107 	 * Bit w&i still works better!
1108 	 */
1109 	tmp |= DMA_WRITE_MAX_128;
1110 #endif
1111 	writel(tmp, &regs->PciState);
1112 
1113 #if 0
1114 	/*
1115 	 * The Host PCI bus controller driver has to set FBB.
1116 	 * If all devices on that PCI bus support FBB, then the controller
1117 	 * can enable FBB support in the Host PCI Bus controller (or on
1118 	 * the PCI-PCI bridge if that applies).
1119 	 * -ggg
1120 	 */
1121 	/*
1122 	 * I have received reports from people having problems when this
1123 	 * bit is enabled.
1124 	 */
1125 	if (!(ap->pci_command & PCI_COMMAND_FAST_BACK)) {
1126 		printk(KERN_INFO "  Enabling PCI Fast Back to Back\n");
1127 		ap->pci_command |= PCI_COMMAND_FAST_BACK;
1128 		pci_write_config_word(pdev, PCI_COMMAND, ap->pci_command);
1129 	}
1130 #endif
1131 
1132 	/*
1133 	 * Configure DMA attributes.
1134 	 */
1135 	if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(64))) {
1136 		ap->pci_using_dac = 1;
1137 	} else if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(32))) {
1138 		ap->pci_using_dac = 0;
1139 	} else {
1140 		ecode = -ENODEV;
1141 		goto init_error;
1142 	}
1143 
1144 	/*
1145 	 * Initialize the generic info block and the command+event rings
1146 	 * and the control blocks for the transmit and receive rings
1147 	 * as they need to be setup once and for all.
1148 	 */
1149 	if (!(info = pci_alloc_consistent(ap->pdev, sizeof(struct ace_info),
1150 					  &ap->info_dma))) {
1151 		ecode = -EAGAIN;
1152 		goto init_error;
1153 	}
1154 	ap->info = info;
1155 
1156 	/*
1157 	 * Get the memory for the skb rings.
1158 	 */
1159 	if (!(ap->skb = kmalloc(sizeof(struct ace_skb), GFP_KERNEL))) {
1160 		ecode = -EAGAIN;
1161 		goto init_error;
1162 	}
1163 
1164 	ecode = request_irq(pdev->irq, ace_interrupt, IRQF_SHARED,
1165 			    DRV_NAME, dev);
1166 	if (ecode) {
1167 		printk(KERN_WARNING "%s: Requested IRQ %d is busy\n",
1168 		       DRV_NAME, pdev->irq);
1169 		goto init_error;
1170 	} else
1171 		dev->irq = pdev->irq;
1172 
1173 #ifdef INDEX_DEBUG
1174 	spin_lock_init(&ap->debug_lock);
1175 	ap->last_tx = ACE_TX_RING_ENTRIES(ap) - 1;
1176 	ap->last_std_rx = 0;
1177 	ap->last_mini_rx = 0;
1178 #endif
1179 
1180 	memset(ap->info, 0, sizeof(struct ace_info));
1181 	memset(ap->skb, 0, sizeof(struct ace_skb));
1182 
1183 	ecode = ace_load_firmware(dev);
1184 	if (ecode)
1185 		goto init_error;
1186 
1187 	ap->fw_running = 0;
1188 
1189 	tmp_ptr = ap->info_dma;
1190 	writel(tmp_ptr >> 32, &regs->InfoPtrHi);
1191 	writel(tmp_ptr & 0xffffffff, &regs->InfoPtrLo);
1192 
1193 	memset(ap->evt_ring, 0, EVT_RING_ENTRIES * sizeof(struct event));
1194 
1195 	set_aceaddr(&info->evt_ctrl.rngptr, ap->evt_ring_dma);
1196 	info->evt_ctrl.flags = 0;
1197 
1198 	*(ap->evt_prd) = 0;
1199 	wmb();
1200 	set_aceaddr(&info->evt_prd_ptr, ap->evt_prd_dma);
1201 	writel(0, &regs->EvtCsm);
1202 
1203 	set_aceaddr(&info->cmd_ctrl.rngptr, 0x100);
1204 	info->cmd_ctrl.flags = 0;
1205 	info->cmd_ctrl.max_len = 0;
1206 
1207 	for (i = 0; i < CMD_RING_ENTRIES; i++)
1208 		writel(0, &regs->CmdRng[i]);
1209 
1210 	writel(0, &regs->CmdPrd);
1211 	writel(0, &regs->CmdCsm);
1212 
1213 	tmp_ptr = ap->info_dma;
1214 	tmp_ptr += (unsigned long) &(((struct ace_info *)0)->s.stats);
1215 	set_aceaddr(&info->stats2_ptr, (dma_addr_t) tmp_ptr);
1216 
1217 	set_aceaddr(&info->rx_std_ctrl.rngptr, ap->rx_ring_base_dma);
1218 	info->rx_std_ctrl.max_len = ACE_STD_BUFSIZE;
1219 	info->rx_std_ctrl.flags =
1220 	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1221 
1222 	memset(ap->rx_std_ring, 0,
1223 	       RX_STD_RING_ENTRIES * sizeof(struct rx_desc));
1224 
1225 	for (i = 0; i < RX_STD_RING_ENTRIES; i++)
1226 		ap->rx_std_ring[i].flags = BD_FLG_TCP_UDP_SUM;
1227 
1228 	ap->rx_std_skbprd = 0;
1229 	atomic_set(&ap->cur_rx_bufs, 0);
1230 
1231 	set_aceaddr(&info->rx_jumbo_ctrl.rngptr,
1232 		    (ap->rx_ring_base_dma +
1233 		     (sizeof(struct rx_desc) * RX_STD_RING_ENTRIES)));
1234 	info->rx_jumbo_ctrl.max_len = 0;
1235 	info->rx_jumbo_ctrl.flags =
1236 	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1237 
1238 	memset(ap->rx_jumbo_ring, 0,
1239 	       RX_JUMBO_RING_ENTRIES * sizeof(struct rx_desc));
1240 
1241 	for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++)
1242 		ap->rx_jumbo_ring[i].flags = BD_FLG_TCP_UDP_SUM | BD_FLG_JUMBO;
1243 
1244 	ap->rx_jumbo_skbprd = 0;
1245 	atomic_set(&ap->cur_jumbo_bufs, 0);
1246 
1247 	memset(ap->rx_mini_ring, 0,
1248 	       RX_MINI_RING_ENTRIES * sizeof(struct rx_desc));
1249 
1250 	if (ap->version >= 2) {
1251 		set_aceaddr(&info->rx_mini_ctrl.rngptr,
1252 			    (ap->rx_ring_base_dma +
1253 			     (sizeof(struct rx_desc) *
1254 			      (RX_STD_RING_ENTRIES +
1255 			       RX_JUMBO_RING_ENTRIES))));
1256 		info->rx_mini_ctrl.max_len = ACE_MINI_SIZE;
1257 		info->rx_mini_ctrl.flags =
1258 		  RCB_FLG_TCP_UDP_SUM|RCB_FLG_NO_PSEUDO_HDR|RCB_FLG_VLAN_ASSIST;
1259 
1260 		for (i = 0; i < RX_MINI_RING_ENTRIES; i++)
1261 			ap->rx_mini_ring[i].flags =
1262 				BD_FLG_TCP_UDP_SUM | BD_FLG_MINI;
1263 	} else {
1264 		set_aceaddr(&info->rx_mini_ctrl.rngptr, 0);
1265 		info->rx_mini_ctrl.flags = RCB_FLG_RNG_DISABLE;
1266 		info->rx_mini_ctrl.max_len = 0;
1267 	}
1268 
1269 	ap->rx_mini_skbprd = 0;
1270 	atomic_set(&ap->cur_mini_bufs, 0);
1271 
1272 	set_aceaddr(&info->rx_return_ctrl.rngptr,
1273 		    (ap->rx_ring_base_dma +
1274 		     (sizeof(struct rx_desc) *
1275 		      (RX_STD_RING_ENTRIES +
1276 		       RX_JUMBO_RING_ENTRIES +
1277 		       RX_MINI_RING_ENTRIES))));
1278 	info->rx_return_ctrl.flags = 0;
1279 	info->rx_return_ctrl.max_len = RX_RETURN_RING_ENTRIES;
1280 
1281 	memset(ap->rx_return_ring, 0,
1282 	       RX_RETURN_RING_ENTRIES * sizeof(struct rx_desc));
1283 
1284 	set_aceaddr(&info->rx_ret_prd_ptr, ap->rx_ret_prd_dma);
1285 	*(ap->rx_ret_prd) = 0;
1286 
1287 	writel(TX_RING_BASE, &regs->WinBase);
1288 
1289 	if (ACE_IS_TIGON_I(ap)) {
1290 		ap->tx_ring = (__force struct tx_desc *) regs->Window;
1291 		for (i = 0; i < (TIGON_I_TX_RING_ENTRIES
1292 				 * sizeof(struct tx_desc)) / sizeof(u32); i++)
1293 			writel(0, (__force void __iomem *)ap->tx_ring  + i * 4);
1294 
1295 		set_aceaddr(&info->tx_ctrl.rngptr, TX_RING_BASE);
1296 	} else {
1297 		memset(ap->tx_ring, 0,
1298 		       MAX_TX_RING_ENTRIES * sizeof(struct tx_desc));
1299 
1300 		set_aceaddr(&info->tx_ctrl.rngptr, ap->tx_ring_dma);
1301 	}
1302 
1303 	info->tx_ctrl.max_len = ACE_TX_RING_ENTRIES(ap);
1304 	tmp = RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1305 
1306 	/*
1307 	 * The Tigon I does not like having the TX ring in host memory ;-(
1308 	 */
1309 	if (!ACE_IS_TIGON_I(ap))
1310 		tmp |= RCB_FLG_TX_HOST_RING;
1311 #if TX_COAL_INTS_ONLY
1312 	tmp |= RCB_FLG_COAL_INT_ONLY;
1313 #endif
1314 	info->tx_ctrl.flags = tmp;
1315 
1316 	set_aceaddr(&info->tx_csm_ptr, ap->tx_csm_dma);
1317 
1318 	/*
1319 	 * Potential item for tuning parameter
1320 	 */
1321 #if 0 /* NO */
1322 	writel(DMA_THRESH_16W, &regs->DmaReadCfg);
1323 	writel(DMA_THRESH_16W, &regs->DmaWriteCfg);
1324 #else
1325 	writel(DMA_THRESH_8W, &regs->DmaReadCfg);
1326 	writel(DMA_THRESH_8W, &regs->DmaWriteCfg);
1327 #endif
1328 
1329 	writel(0, &regs->MaskInt);
1330 	writel(1, &regs->IfIdx);
1331 #if 0
1332 	/*
1333 	 * McKinley boxes do not like us fiddling with AssistState
1334 	 * this early
1335 	 */
1336 	writel(1, &regs->AssistState);
1337 #endif
1338 
1339 	writel(DEF_STAT, &regs->TuneStatTicks);
1340 	writel(DEF_TRACE, &regs->TuneTrace);
1341 
1342 	ace_set_rxtx_parms(dev, 0);
1343 
1344 	if (board_idx == BOARD_IDX_OVERFLOW) {
1345 		printk(KERN_WARNING "%s: more than %i NICs detected, "
1346 		       "ignoring module parameters!\n",
1347 		       ap->name, ACE_MAX_MOD_PARMS);
1348 	} else if (board_idx >= 0) {
1349 		if (tx_coal_tick[board_idx])
1350 			writel(tx_coal_tick[board_idx],
1351 			       &regs->TuneTxCoalTicks);
1352 		if (max_tx_desc[board_idx])
1353 			writel(max_tx_desc[board_idx], &regs->TuneMaxTxDesc);
1354 
1355 		if (rx_coal_tick[board_idx])
1356 			writel(rx_coal_tick[board_idx],
1357 			       &regs->TuneRxCoalTicks);
1358 		if (max_rx_desc[board_idx])
1359 			writel(max_rx_desc[board_idx], &regs->TuneMaxRxDesc);
1360 
1361 		if (trace[board_idx])
1362 			writel(trace[board_idx], &regs->TuneTrace);
1363 
1364 		if ((tx_ratio[board_idx] > 0) && (tx_ratio[board_idx] < 64))
1365 			writel(tx_ratio[board_idx], &regs->TxBufRat);
1366 	}
1367 
1368 	/*
1369 	 * Default link parameters
1370 	 */
1371 	tmp = LNK_ENABLE | LNK_FULL_DUPLEX | LNK_1000MB | LNK_100MB |
1372 		LNK_10MB | LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL | LNK_NEGOTIATE;
1373 	if(ap->version >= 2)
1374 		tmp |= LNK_TX_FLOW_CTL_Y;
1375 
1376 	/*
1377 	 * Override link default parameters
1378 	 */
1379 	if ((board_idx >= 0) && link_state[board_idx]) {
1380 		int option = link_state[board_idx];
1381 
1382 		tmp = LNK_ENABLE;
1383 
1384 		if (option & 0x01) {
1385 			printk(KERN_INFO "%s: Setting half duplex link\n",
1386 			       ap->name);
1387 			tmp &= ~LNK_FULL_DUPLEX;
1388 		}
1389 		if (option & 0x02)
1390 			tmp &= ~LNK_NEGOTIATE;
1391 		if (option & 0x10)
1392 			tmp |= LNK_10MB;
1393 		if (option & 0x20)
1394 			tmp |= LNK_100MB;
1395 		if (option & 0x40)
1396 			tmp |= LNK_1000MB;
1397 		if ((option & 0x70) == 0) {
1398 			printk(KERN_WARNING "%s: No media speed specified, "
1399 			       "forcing auto negotiation\n", ap->name);
1400 			tmp |= LNK_NEGOTIATE | LNK_1000MB |
1401 				LNK_100MB | LNK_10MB;
1402 		}
1403 		if ((option & 0x100) == 0)
1404 			tmp |= LNK_NEG_FCTL;
1405 		else
1406 			printk(KERN_INFO "%s: Disabling flow control "
1407 			       "negotiation\n", ap->name);
1408 		if (option & 0x200)
1409 			tmp |= LNK_RX_FLOW_CTL_Y;
1410 		if ((option & 0x400) && (ap->version >= 2)) {
1411 			printk(KERN_INFO "%s: Enabling TX flow control\n",
1412 			       ap->name);
1413 			tmp |= LNK_TX_FLOW_CTL_Y;
1414 		}
1415 	}
1416 
1417 	ap->link = tmp;
1418 	writel(tmp, &regs->TuneLink);
1419 	if (ap->version >= 2)
1420 		writel(tmp, &regs->TuneFastLink);
1421 
1422 	writel(ap->firmware_start, &regs->Pc);
1423 
1424 	writel(0, &regs->Mb0Lo);
1425 
1426 	/*
1427 	 * Set tx_csm before we start receiving interrupts, otherwise
1428 	 * the interrupt handler might think it is supposed to process
1429 	 * tx ints before we are up and running, which may cause a null
1430 	 * pointer access in the int handler.
1431 	 */
1432 	ap->cur_rx = 0;
1433 	ap->tx_prd = *(ap->tx_csm) = ap->tx_ret_csm = 0;
1434 
1435 	wmb();
1436 	ace_set_txprd(regs, ap, 0);
1437 	writel(0, &regs->RxRetCsm);
1438 
1439 	/*
1440 	 * Enable DMA engine now.
1441 	 * If we do this sooner, Mckinley box pukes.
1442 	 * I assume it's because Tigon II DMA engine wants to check
1443 	 * *something* even before the CPU is started.
1444 	 */
1445 	writel(1, &regs->AssistState);  /* enable DMA */
1446 
1447 	/*
1448 	 * Start the NIC CPU
1449 	 */
1450 	writel(readl(&regs->CpuCtrl) & ~(CPU_HALT|CPU_TRACE), &regs->CpuCtrl);
1451 	readl(&regs->CpuCtrl);
1452 
1453 	/*
1454 	 * Wait for the firmware to spin up - max 3 seconds.
1455 	 */
1456 	myjif = jiffies + 3 * HZ;
1457 	while (time_before(jiffies, myjif) && !ap->fw_running)
1458 		cpu_relax();
1459 
1460 	if (!ap->fw_running) {
1461 		printk(KERN_ERR "%s: Firmware NOT running!\n", ap->name);
1462 
1463 		ace_dump_trace(ap);
1464 		writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
1465 		readl(&regs->CpuCtrl);
1466 
1467 		/* aman@sgi.com - account for badly behaving firmware/NIC:
1468 		 * - have observed that the NIC may continue to generate
1469 		 *   interrupts for some reason; attempt to stop it - halt
1470 		 *   second CPU for Tigon II cards, and also clear Mb0
1471 		 * - if we're a module, we'll fail to load if this was
1472 		 *   the only GbE card in the system => if the kernel does
1473 		 *   see an interrupt from the NIC, code to handle it is
1474 		 *   gone and OOps! - so free_irq also
1475 		 */
1476 		if (ap->version >= 2)
1477 			writel(readl(&regs->CpuBCtrl) | CPU_HALT,
1478 			       &regs->CpuBCtrl);
1479 		writel(0, &regs->Mb0Lo);
1480 		readl(&regs->Mb0Lo);
1481 
1482 		ecode = -EBUSY;
1483 		goto init_error;
1484 	}
1485 
1486 	/*
1487 	 * We load the ring here as there seem to be no way to tell the
1488 	 * firmware to wipe the ring without re-initializing it.
1489 	 */
1490 	if (!test_and_set_bit(0, &ap->std_refill_busy))
1491 		ace_load_std_rx_ring(dev, RX_RING_SIZE);
1492 	else
1493 		printk(KERN_ERR "%s: Someone is busy refilling the RX ring\n",
1494 		       ap->name);
1495 	if (ap->version >= 2) {
1496 		if (!test_and_set_bit(0, &ap->mini_refill_busy))
1497 			ace_load_mini_rx_ring(dev, RX_MINI_SIZE);
1498 		else
1499 			printk(KERN_ERR "%s: Someone is busy refilling "
1500 			       "the RX mini ring\n", ap->name);
1501 	}
1502 	return 0;
1503 
1504  init_error:
1505 	ace_init_cleanup(dev);
1506 	return ecode;
1507 }
1508 
1509 
1510 static void ace_set_rxtx_parms(struct net_device *dev, int jumbo)
1511 {
1512 	struct ace_private *ap = netdev_priv(dev);
1513 	struct ace_regs __iomem *regs = ap->regs;
1514 	int board_idx = ap->board_idx;
1515 
1516 	if (board_idx >= 0) {
1517 		if (!jumbo) {
1518 			if (!tx_coal_tick[board_idx])
1519 				writel(DEF_TX_COAL, &regs->TuneTxCoalTicks);
1520 			if (!max_tx_desc[board_idx])
1521 				writel(DEF_TX_MAX_DESC, &regs->TuneMaxTxDesc);
1522 			if (!rx_coal_tick[board_idx])
1523 				writel(DEF_RX_COAL, &regs->TuneRxCoalTicks);
1524 			if (!max_rx_desc[board_idx])
1525 				writel(DEF_RX_MAX_DESC, &regs->TuneMaxRxDesc);
1526 			if (!tx_ratio[board_idx])
1527 				writel(DEF_TX_RATIO, &regs->TxBufRat);
1528 		} else {
1529 			if (!tx_coal_tick[board_idx])
1530 				writel(DEF_JUMBO_TX_COAL,
1531 				       &regs->TuneTxCoalTicks);
1532 			if (!max_tx_desc[board_idx])
1533 				writel(DEF_JUMBO_TX_MAX_DESC,
1534 				       &regs->TuneMaxTxDesc);
1535 			if (!rx_coal_tick[board_idx])
1536 				writel(DEF_JUMBO_RX_COAL,
1537 				       &regs->TuneRxCoalTicks);
1538 			if (!max_rx_desc[board_idx])
1539 				writel(DEF_JUMBO_RX_MAX_DESC,
1540 				       &regs->TuneMaxRxDesc);
1541 			if (!tx_ratio[board_idx])
1542 				writel(DEF_JUMBO_TX_RATIO, &regs->TxBufRat);
1543 		}
1544 	}
1545 }
1546 
1547 
1548 static void ace_watchdog(struct net_device *data)
1549 {
1550 	struct net_device *dev = data;
1551 	struct ace_private *ap = netdev_priv(dev);
1552 	struct ace_regs __iomem *regs = ap->regs;
1553 
1554 	/*
1555 	 * We haven't received a stats update event for more than 2.5
1556 	 * seconds and there is data in the transmit queue, thus we
1557 	 * assume the card is stuck.
1558 	 */
1559 	if (*ap->tx_csm != ap->tx_ret_csm) {
1560 		printk(KERN_WARNING "%s: Transmitter is stuck, %08x\n",
1561 		       dev->name, (unsigned int)readl(&regs->HostCtrl));
1562 		/* This can happen due to ieee flow control. */
1563 	} else {
1564 		printk(KERN_DEBUG "%s: BUG... transmitter died. Kicking it.\n",
1565 		       dev->name);
1566 #if 0
1567 		netif_wake_queue(dev);
1568 #endif
1569 	}
1570 }
1571 
1572 
1573 static void ace_tasklet(unsigned long arg)
1574 {
1575 	struct net_device *dev = (struct net_device *) arg;
1576 	struct ace_private *ap = netdev_priv(dev);
1577 	int cur_size;
1578 
1579 	cur_size = atomic_read(&ap->cur_rx_bufs);
1580 	if ((cur_size < RX_LOW_STD_THRES) &&
1581 	    !test_and_set_bit(0, &ap->std_refill_busy)) {
1582 #ifdef DEBUG
1583 		printk("refilling buffers (current %i)\n", cur_size);
1584 #endif
1585 		ace_load_std_rx_ring(dev, RX_RING_SIZE - cur_size);
1586 	}
1587 
1588 	if (ap->version >= 2) {
1589 		cur_size = atomic_read(&ap->cur_mini_bufs);
1590 		if ((cur_size < RX_LOW_MINI_THRES) &&
1591 		    !test_and_set_bit(0, &ap->mini_refill_busy)) {
1592 #ifdef DEBUG
1593 			printk("refilling mini buffers (current %i)\n",
1594 			       cur_size);
1595 #endif
1596 			ace_load_mini_rx_ring(dev, RX_MINI_SIZE - cur_size);
1597 		}
1598 	}
1599 
1600 	cur_size = atomic_read(&ap->cur_jumbo_bufs);
1601 	if (ap->jumbo && (cur_size < RX_LOW_JUMBO_THRES) &&
1602 	    !test_and_set_bit(0, &ap->jumbo_refill_busy)) {
1603 #ifdef DEBUG
1604 		printk("refilling jumbo buffers (current %i)\n", cur_size);
1605 #endif
1606 		ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE - cur_size);
1607 	}
1608 	ap->tasklet_pending = 0;
1609 }
1610 
1611 
1612 /*
1613  * Copy the contents of the NIC's trace buffer to kernel memory.
1614  */
1615 static void ace_dump_trace(struct ace_private *ap)
1616 {
1617 #if 0
1618 	if (!ap->trace_buf)
1619 		if (!(ap->trace_buf = kmalloc(ACE_TRACE_SIZE, GFP_KERNEL)))
1620 		    return;
1621 #endif
1622 }
1623 
1624 
1625 /*
1626  * Load the standard rx ring.
1627  *
1628  * Loading rings is safe without holding the spin lock since this is
1629  * done only before the device is enabled, thus no interrupts are
1630  * generated and by the interrupt handler/tasklet handler.
1631  */
1632 static void ace_load_std_rx_ring(struct net_device *dev, int nr_bufs)
1633 {
1634 	struct ace_private *ap = netdev_priv(dev);
1635 	struct ace_regs __iomem *regs = ap->regs;
1636 	short i, idx;
1637 
1638 
1639 	prefetchw(&ap->cur_rx_bufs);
1640 
1641 	idx = ap->rx_std_skbprd;
1642 
1643 	for (i = 0; i < nr_bufs; i++) {
1644 		struct sk_buff *skb;
1645 		struct rx_desc *rd;
1646 		dma_addr_t mapping;
1647 
1648 		skb = netdev_alloc_skb_ip_align(dev, ACE_STD_BUFSIZE);
1649 		if (!skb)
1650 			break;
1651 
1652 		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1653 				       offset_in_page(skb->data),
1654 				       ACE_STD_BUFSIZE,
1655 				       PCI_DMA_FROMDEVICE);
1656 		ap->skb->rx_std_skbuff[idx].skb = skb;
1657 		dma_unmap_addr_set(&ap->skb->rx_std_skbuff[idx],
1658 				   mapping, mapping);
1659 
1660 		rd = &ap->rx_std_ring[idx];
1661 		set_aceaddr(&rd->addr, mapping);
1662 		rd->size = ACE_STD_BUFSIZE;
1663 		rd->idx = idx;
1664 		idx = (idx + 1) % RX_STD_RING_ENTRIES;
1665 	}
1666 
1667 	if (!i)
1668 		goto error_out;
1669 
1670 	atomic_add(i, &ap->cur_rx_bufs);
1671 	ap->rx_std_skbprd = idx;
1672 
1673 	if (ACE_IS_TIGON_I(ap)) {
1674 		struct cmd cmd;
1675 		cmd.evt = C_SET_RX_PRD_IDX;
1676 		cmd.code = 0;
1677 		cmd.idx = ap->rx_std_skbprd;
1678 		ace_issue_cmd(regs, &cmd);
1679 	} else {
1680 		writel(idx, &regs->RxStdPrd);
1681 		wmb();
1682 	}
1683 
1684  out:
1685 	clear_bit(0, &ap->std_refill_busy);
1686 	return;
1687 
1688  error_out:
1689 	printk(KERN_INFO "Out of memory when allocating "
1690 	       "standard receive buffers\n");
1691 	goto out;
1692 }
1693 
1694 
1695 static void ace_load_mini_rx_ring(struct net_device *dev, int nr_bufs)
1696 {
1697 	struct ace_private *ap = netdev_priv(dev);
1698 	struct ace_regs __iomem *regs = ap->regs;
1699 	short i, idx;
1700 
1701 	prefetchw(&ap->cur_mini_bufs);
1702 
1703 	idx = ap->rx_mini_skbprd;
1704 	for (i = 0; i < nr_bufs; i++) {
1705 		struct sk_buff *skb;
1706 		struct rx_desc *rd;
1707 		dma_addr_t mapping;
1708 
1709 		skb = netdev_alloc_skb_ip_align(dev, ACE_MINI_BUFSIZE);
1710 		if (!skb)
1711 			break;
1712 
1713 		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1714 				       offset_in_page(skb->data),
1715 				       ACE_MINI_BUFSIZE,
1716 				       PCI_DMA_FROMDEVICE);
1717 		ap->skb->rx_mini_skbuff[idx].skb = skb;
1718 		dma_unmap_addr_set(&ap->skb->rx_mini_skbuff[idx],
1719 				   mapping, mapping);
1720 
1721 		rd = &ap->rx_mini_ring[idx];
1722 		set_aceaddr(&rd->addr, mapping);
1723 		rd->size = ACE_MINI_BUFSIZE;
1724 		rd->idx = idx;
1725 		idx = (idx + 1) % RX_MINI_RING_ENTRIES;
1726 	}
1727 
1728 	if (!i)
1729 		goto error_out;
1730 
1731 	atomic_add(i, &ap->cur_mini_bufs);
1732 
1733 	ap->rx_mini_skbprd = idx;
1734 
1735 	writel(idx, &regs->RxMiniPrd);
1736 	wmb();
1737 
1738  out:
1739 	clear_bit(0, &ap->mini_refill_busy);
1740 	return;
1741  error_out:
1742 	printk(KERN_INFO "Out of memory when allocating "
1743 	       "mini receive buffers\n");
1744 	goto out;
1745 }
1746 
1747 
1748 /*
1749  * Load the jumbo rx ring, this may happen at any time if the MTU
1750  * is changed to a value > 1500.
1751  */
1752 static void ace_load_jumbo_rx_ring(struct net_device *dev, int nr_bufs)
1753 {
1754 	struct ace_private *ap = netdev_priv(dev);
1755 	struct ace_regs __iomem *regs = ap->regs;
1756 	short i, idx;
1757 
1758 	idx = ap->rx_jumbo_skbprd;
1759 
1760 	for (i = 0; i < nr_bufs; i++) {
1761 		struct sk_buff *skb;
1762 		struct rx_desc *rd;
1763 		dma_addr_t mapping;
1764 
1765 		skb = netdev_alloc_skb_ip_align(dev, ACE_JUMBO_BUFSIZE);
1766 		if (!skb)
1767 			break;
1768 
1769 		mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1770 				       offset_in_page(skb->data),
1771 				       ACE_JUMBO_BUFSIZE,
1772 				       PCI_DMA_FROMDEVICE);
1773 		ap->skb->rx_jumbo_skbuff[idx].skb = skb;
1774 		dma_unmap_addr_set(&ap->skb->rx_jumbo_skbuff[idx],
1775 				   mapping, mapping);
1776 
1777 		rd = &ap->rx_jumbo_ring[idx];
1778 		set_aceaddr(&rd->addr, mapping);
1779 		rd->size = ACE_JUMBO_BUFSIZE;
1780 		rd->idx = idx;
1781 		idx = (idx + 1) % RX_JUMBO_RING_ENTRIES;
1782 	}
1783 
1784 	if (!i)
1785 		goto error_out;
1786 
1787 	atomic_add(i, &ap->cur_jumbo_bufs);
1788 	ap->rx_jumbo_skbprd = idx;
1789 
1790 	if (ACE_IS_TIGON_I(ap)) {
1791 		struct cmd cmd;
1792 		cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1793 		cmd.code = 0;
1794 		cmd.idx = ap->rx_jumbo_skbprd;
1795 		ace_issue_cmd(regs, &cmd);
1796 	} else {
1797 		writel(idx, &regs->RxJumboPrd);
1798 		wmb();
1799 	}
1800 
1801  out:
1802 	clear_bit(0, &ap->jumbo_refill_busy);
1803 	return;
1804  error_out:
1805 	if (net_ratelimit())
1806 		printk(KERN_INFO "Out of memory when allocating "
1807 		       "jumbo receive buffers\n");
1808 	goto out;
1809 }
1810 
1811 
1812 /*
1813  * All events are considered to be slow (RX/TX ints do not generate
1814  * events) and are handled here, outside the main interrupt handler,
1815  * to reduce the size of the handler.
1816  */
1817 static u32 ace_handle_event(struct net_device *dev, u32 evtcsm, u32 evtprd)
1818 {
1819 	struct ace_private *ap;
1820 
1821 	ap = netdev_priv(dev);
1822 
1823 	while (evtcsm != evtprd) {
1824 		switch (ap->evt_ring[evtcsm].evt) {
1825 		case E_FW_RUNNING:
1826 			printk(KERN_INFO "%s: Firmware up and running\n",
1827 			       ap->name);
1828 			ap->fw_running = 1;
1829 			wmb();
1830 			break;
1831 		case E_STATS_UPDATED:
1832 			break;
1833 		case E_LNK_STATE:
1834 		{
1835 			u16 code = ap->evt_ring[evtcsm].code;
1836 			switch (code) {
1837 			case E_C_LINK_UP:
1838 			{
1839 				u32 state = readl(&ap->regs->GigLnkState);
1840 				printk(KERN_WARNING "%s: Optical link UP "
1841 				       "(%s Duplex, Flow Control: %s%s)\n",
1842 				       ap->name,
1843 				       state & LNK_FULL_DUPLEX ? "Full":"Half",
1844 				       state & LNK_TX_FLOW_CTL_Y ? "TX " : "",
1845 				       state & LNK_RX_FLOW_CTL_Y ? "RX" : "");
1846 				break;
1847 			}
1848 			case E_C_LINK_DOWN:
1849 				printk(KERN_WARNING "%s: Optical link DOWN\n",
1850 				       ap->name);
1851 				break;
1852 			case E_C_LINK_10_100:
1853 				printk(KERN_WARNING "%s: 10/100BaseT link "
1854 				       "UP\n", ap->name);
1855 				break;
1856 			default:
1857 				printk(KERN_ERR "%s: Unknown optical link "
1858 				       "state %02x\n", ap->name, code);
1859 			}
1860 			break;
1861 		}
1862 		case E_ERROR:
1863 			switch(ap->evt_ring[evtcsm].code) {
1864 			case E_C_ERR_INVAL_CMD:
1865 				printk(KERN_ERR "%s: invalid command error\n",
1866 				       ap->name);
1867 				break;
1868 			case E_C_ERR_UNIMP_CMD:
1869 				printk(KERN_ERR "%s: unimplemented command "
1870 				       "error\n", ap->name);
1871 				break;
1872 			case E_C_ERR_BAD_CFG:
1873 				printk(KERN_ERR "%s: bad config error\n",
1874 				       ap->name);
1875 				break;
1876 			default:
1877 				printk(KERN_ERR "%s: unknown error %02x\n",
1878 				       ap->name, ap->evt_ring[evtcsm].code);
1879 			}
1880 			break;
1881 		case E_RESET_JUMBO_RNG:
1882 		{
1883 			int i;
1884 			for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
1885 				if (ap->skb->rx_jumbo_skbuff[i].skb) {
1886 					ap->rx_jumbo_ring[i].size = 0;
1887 					set_aceaddr(&ap->rx_jumbo_ring[i].addr, 0);
1888 					dev_kfree_skb(ap->skb->rx_jumbo_skbuff[i].skb);
1889 					ap->skb->rx_jumbo_skbuff[i].skb = NULL;
1890 				}
1891 			}
1892 
1893  			if (ACE_IS_TIGON_I(ap)) {
1894  				struct cmd cmd;
1895  				cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1896  				cmd.code = 0;
1897  				cmd.idx = 0;
1898  				ace_issue_cmd(ap->regs, &cmd);
1899  			} else {
1900  				writel(0, &((ap->regs)->RxJumboPrd));
1901  				wmb();
1902  			}
1903 
1904 			ap->jumbo = 0;
1905 			ap->rx_jumbo_skbprd = 0;
1906 			printk(KERN_INFO "%s: Jumbo ring flushed\n",
1907 			       ap->name);
1908 			clear_bit(0, &ap->jumbo_refill_busy);
1909 			break;
1910 		}
1911 		default:
1912 			printk(KERN_ERR "%s: Unhandled event 0x%02x\n",
1913 			       ap->name, ap->evt_ring[evtcsm].evt);
1914 		}
1915 		evtcsm = (evtcsm + 1) % EVT_RING_ENTRIES;
1916 	}
1917 
1918 	return evtcsm;
1919 }
1920 
1921 
1922 static void ace_rx_int(struct net_device *dev, u32 rxretprd, u32 rxretcsm)
1923 {
1924 	struct ace_private *ap = netdev_priv(dev);
1925 	u32 idx;
1926 	int mini_count = 0, std_count = 0;
1927 
1928 	idx = rxretcsm;
1929 
1930 	prefetchw(&ap->cur_rx_bufs);
1931 	prefetchw(&ap->cur_mini_bufs);
1932 
1933 	while (idx != rxretprd) {
1934 		struct ring_info *rip;
1935 		struct sk_buff *skb;
1936 		struct rx_desc *retdesc;
1937 		u32 skbidx;
1938 		int bd_flags, desc_type, mapsize;
1939 		u16 csum;
1940 
1941 
1942 		/* make sure the rx descriptor isn't read before rxretprd */
1943 		if (idx == rxretcsm)
1944 			rmb();
1945 
1946 		retdesc = &ap->rx_return_ring[idx];
1947 		skbidx = retdesc->idx;
1948 		bd_flags = retdesc->flags;
1949 		desc_type = bd_flags & (BD_FLG_JUMBO | BD_FLG_MINI);
1950 
1951 		switch(desc_type) {
1952 			/*
1953 			 * Normal frames do not have any flags set
1954 			 *
1955 			 * Mini and normal frames arrive frequently,
1956 			 * so use a local counter to avoid doing
1957 			 * atomic operations for each packet arriving.
1958 			 */
1959 		case 0:
1960 			rip = &ap->skb->rx_std_skbuff[skbidx];
1961 			mapsize = ACE_STD_BUFSIZE;
1962 			std_count++;
1963 			break;
1964 		case BD_FLG_JUMBO:
1965 			rip = &ap->skb->rx_jumbo_skbuff[skbidx];
1966 			mapsize = ACE_JUMBO_BUFSIZE;
1967 			atomic_dec(&ap->cur_jumbo_bufs);
1968 			break;
1969 		case BD_FLG_MINI:
1970 			rip = &ap->skb->rx_mini_skbuff[skbidx];
1971 			mapsize = ACE_MINI_BUFSIZE;
1972 			mini_count++;
1973 			break;
1974 		default:
1975 			printk(KERN_INFO "%s: unknown frame type (0x%02x) "
1976 			       "returned by NIC\n", dev->name,
1977 			       retdesc->flags);
1978 			goto error;
1979 		}
1980 
1981 		skb = rip->skb;
1982 		rip->skb = NULL;
1983 		pci_unmap_page(ap->pdev,
1984 			       dma_unmap_addr(rip, mapping),
1985 			       mapsize,
1986 			       PCI_DMA_FROMDEVICE);
1987 		skb_put(skb, retdesc->size);
1988 
1989 		/*
1990 		 * Fly baby, fly!
1991 		 */
1992 		csum = retdesc->tcp_udp_csum;
1993 
1994 		skb->protocol = eth_type_trans(skb, dev);
1995 
1996 		/*
1997 		 * Instead of forcing the poor tigon mips cpu to calculate
1998 		 * pseudo hdr checksum, we do this ourselves.
1999 		 */
2000 		if (bd_flags & BD_FLG_TCP_UDP_SUM) {
2001 			skb->csum = htons(csum);
2002 			skb->ip_summed = CHECKSUM_COMPLETE;
2003 		} else {
2004 			skb_checksum_none_assert(skb);
2005 		}
2006 
2007 		/* send it up */
2008 		if ((bd_flags & BD_FLG_VLAN_TAG))
2009 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), retdesc->vlan);
2010 		netif_rx(skb);
2011 
2012 		dev->stats.rx_packets++;
2013 		dev->stats.rx_bytes += retdesc->size;
2014 
2015 		idx = (idx + 1) % RX_RETURN_RING_ENTRIES;
2016 	}
2017 
2018 	atomic_sub(std_count, &ap->cur_rx_bufs);
2019 	if (!ACE_IS_TIGON_I(ap))
2020 		atomic_sub(mini_count, &ap->cur_mini_bufs);
2021 
2022  out:
2023 	/*
2024 	 * According to the documentation RxRetCsm is obsolete with
2025 	 * the 12.3.x Firmware - my Tigon I NICs seem to disagree!
2026 	 */
2027 	if (ACE_IS_TIGON_I(ap)) {
2028 		writel(idx, &ap->regs->RxRetCsm);
2029 	}
2030 	ap->cur_rx = idx;
2031 
2032 	return;
2033  error:
2034 	idx = rxretprd;
2035 	goto out;
2036 }
2037 
2038 
2039 static inline void ace_tx_int(struct net_device *dev,
2040 			      u32 txcsm, u32 idx)
2041 {
2042 	struct ace_private *ap = netdev_priv(dev);
2043 
2044 	do {
2045 		struct sk_buff *skb;
2046 		struct tx_ring_info *info;
2047 
2048 		info = ap->skb->tx_skbuff + idx;
2049 		skb = info->skb;
2050 
2051 		if (dma_unmap_len(info, maplen)) {
2052 			pci_unmap_page(ap->pdev, dma_unmap_addr(info, mapping),
2053 				       dma_unmap_len(info, maplen),
2054 				       PCI_DMA_TODEVICE);
2055 			dma_unmap_len_set(info, maplen, 0);
2056 		}
2057 
2058 		if (skb) {
2059 			dev->stats.tx_packets++;
2060 			dev->stats.tx_bytes += skb->len;
2061 			dev_kfree_skb_irq(skb);
2062 			info->skb = NULL;
2063 		}
2064 
2065 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2066 	} while (idx != txcsm);
2067 
2068 	if (netif_queue_stopped(dev))
2069 		netif_wake_queue(dev);
2070 
2071 	wmb();
2072 	ap->tx_ret_csm = txcsm;
2073 
2074 	/* So... tx_ret_csm is advanced _after_ check for device wakeup.
2075 	 *
2076 	 * We could try to make it before. In this case we would get
2077 	 * the following race condition: hard_start_xmit on other cpu
2078 	 * enters after we advanced tx_ret_csm and fills space,
2079 	 * which we have just freed, so that we make illegal device wakeup.
2080 	 * There is no good way to workaround this (at entry
2081 	 * to ace_start_xmit detects this condition and prevents
2082 	 * ring corruption, but it is not a good workaround.)
2083 	 *
2084 	 * When tx_ret_csm is advanced after, we wake up device _only_
2085 	 * if we really have some space in ring (though the core doing
2086 	 * hard_start_xmit can see full ring for some period and has to
2087 	 * synchronize.) Superb.
2088 	 * BUT! We get another subtle race condition. hard_start_xmit
2089 	 * may think that ring is full between wakeup and advancing
2090 	 * tx_ret_csm and will stop device instantly! It is not so bad.
2091 	 * We are guaranteed that there is something in ring, so that
2092 	 * the next irq will resume transmission. To speedup this we could
2093 	 * mark descriptor, which closes ring with BD_FLG_COAL_NOW
2094 	 * (see ace_start_xmit).
2095 	 *
2096 	 * Well, this dilemma exists in all lock-free devices.
2097 	 * We, following scheme used in drivers by Donald Becker,
2098 	 * select the least dangerous.
2099 	 *							--ANK
2100 	 */
2101 }
2102 
2103 
2104 static irqreturn_t ace_interrupt(int irq, void *dev_id)
2105 {
2106 	struct net_device *dev = (struct net_device *)dev_id;
2107 	struct ace_private *ap = netdev_priv(dev);
2108 	struct ace_regs __iomem *regs = ap->regs;
2109 	u32 idx;
2110 	u32 txcsm, rxretcsm, rxretprd;
2111 	u32 evtcsm, evtprd;
2112 
2113 	/*
2114 	 * In case of PCI shared interrupts or spurious interrupts,
2115 	 * we want to make sure it is actually our interrupt before
2116 	 * spending any time in here.
2117 	 */
2118 	if (!(readl(&regs->HostCtrl) & IN_INT))
2119 		return IRQ_NONE;
2120 
2121 	/*
2122 	 * ACK intr now. Otherwise we will lose updates to rx_ret_prd,
2123 	 * which happened _after_ rxretprd = *ap->rx_ret_prd; but before
2124 	 * writel(0, &regs->Mb0Lo).
2125 	 *
2126 	 * "IRQ avoidance" recommended in docs applies to IRQs served
2127 	 * threads and it is wrong even for that case.
2128 	 */
2129 	writel(0, &regs->Mb0Lo);
2130 	readl(&regs->Mb0Lo);
2131 
2132 	/*
2133 	 * There is no conflict between transmit handling in
2134 	 * start_xmit and receive processing, thus there is no reason
2135 	 * to take a spin lock for RX handling. Wait until we start
2136 	 * working on the other stuff - hey we don't need a spin lock
2137 	 * anymore.
2138 	 */
2139 	rxretprd = *ap->rx_ret_prd;
2140 	rxretcsm = ap->cur_rx;
2141 
2142 	if (rxretprd != rxretcsm)
2143 		ace_rx_int(dev, rxretprd, rxretcsm);
2144 
2145 	txcsm = *ap->tx_csm;
2146 	idx = ap->tx_ret_csm;
2147 
2148 	if (txcsm != idx) {
2149 		/*
2150 		 * If each skb takes only one descriptor this check degenerates
2151 		 * to identity, because new space has just been opened.
2152 		 * But if skbs are fragmented we must check that this index
2153 		 * update releases enough of space, otherwise we just
2154 		 * wait for device to make more work.
2155 		 */
2156 		if (!tx_ring_full(ap, txcsm, ap->tx_prd))
2157 			ace_tx_int(dev, txcsm, idx);
2158 	}
2159 
2160 	evtcsm = readl(&regs->EvtCsm);
2161 	evtprd = *ap->evt_prd;
2162 
2163 	if (evtcsm != evtprd) {
2164 		evtcsm = ace_handle_event(dev, evtcsm, evtprd);
2165 		writel(evtcsm, &regs->EvtCsm);
2166 	}
2167 
2168 	/*
2169 	 * This has to go last in the interrupt handler and run with
2170 	 * the spin lock released ... what lock?
2171 	 */
2172 	if (netif_running(dev)) {
2173 		int cur_size;
2174 		int run_tasklet = 0;
2175 
2176 		cur_size = atomic_read(&ap->cur_rx_bufs);
2177 		if (cur_size < RX_LOW_STD_THRES) {
2178 			if ((cur_size < RX_PANIC_STD_THRES) &&
2179 			    !test_and_set_bit(0, &ap->std_refill_busy)) {
2180 #ifdef DEBUG
2181 				printk("low on std buffers %i\n", cur_size);
2182 #endif
2183 				ace_load_std_rx_ring(dev,
2184 						     RX_RING_SIZE - cur_size);
2185 			} else
2186 				run_tasklet = 1;
2187 		}
2188 
2189 		if (!ACE_IS_TIGON_I(ap)) {
2190 			cur_size = atomic_read(&ap->cur_mini_bufs);
2191 			if (cur_size < RX_LOW_MINI_THRES) {
2192 				if ((cur_size < RX_PANIC_MINI_THRES) &&
2193 				    !test_and_set_bit(0,
2194 						      &ap->mini_refill_busy)) {
2195 #ifdef DEBUG
2196 					printk("low on mini buffers %i\n",
2197 					       cur_size);
2198 #endif
2199 					ace_load_mini_rx_ring(dev,
2200 							      RX_MINI_SIZE - cur_size);
2201 				} else
2202 					run_tasklet = 1;
2203 			}
2204 		}
2205 
2206 		if (ap->jumbo) {
2207 			cur_size = atomic_read(&ap->cur_jumbo_bufs);
2208 			if (cur_size < RX_LOW_JUMBO_THRES) {
2209 				if ((cur_size < RX_PANIC_JUMBO_THRES) &&
2210 				    !test_and_set_bit(0,
2211 						      &ap->jumbo_refill_busy)){
2212 #ifdef DEBUG
2213 					printk("low on jumbo buffers %i\n",
2214 					       cur_size);
2215 #endif
2216 					ace_load_jumbo_rx_ring(dev,
2217 							       RX_JUMBO_SIZE - cur_size);
2218 				} else
2219 					run_tasklet = 1;
2220 			}
2221 		}
2222 		if (run_tasklet && !ap->tasklet_pending) {
2223 			ap->tasklet_pending = 1;
2224 			tasklet_schedule(&ap->ace_tasklet);
2225 		}
2226 	}
2227 
2228 	return IRQ_HANDLED;
2229 }
2230 
2231 static int ace_open(struct net_device *dev)
2232 {
2233 	struct ace_private *ap = netdev_priv(dev);
2234 	struct ace_regs __iomem *regs = ap->regs;
2235 	struct cmd cmd;
2236 
2237 	if (!(ap->fw_running)) {
2238 		printk(KERN_WARNING "%s: Firmware not running!\n", dev->name);
2239 		return -EBUSY;
2240 	}
2241 
2242 	writel(dev->mtu + ETH_HLEN + 4, &regs->IfMtu);
2243 
2244 	cmd.evt = C_CLEAR_STATS;
2245 	cmd.code = 0;
2246 	cmd.idx = 0;
2247 	ace_issue_cmd(regs, &cmd);
2248 
2249 	cmd.evt = C_HOST_STATE;
2250 	cmd.code = C_C_STACK_UP;
2251 	cmd.idx = 0;
2252 	ace_issue_cmd(regs, &cmd);
2253 
2254 	if (ap->jumbo &&
2255 	    !test_and_set_bit(0, &ap->jumbo_refill_busy))
2256 		ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE);
2257 
2258 	if (dev->flags & IFF_PROMISC) {
2259 		cmd.evt = C_SET_PROMISC_MODE;
2260 		cmd.code = C_C_PROMISC_ENABLE;
2261 		cmd.idx = 0;
2262 		ace_issue_cmd(regs, &cmd);
2263 
2264 		ap->promisc = 1;
2265 	}else
2266 		ap->promisc = 0;
2267 	ap->mcast_all = 0;
2268 
2269 #if 0
2270 	cmd.evt = C_LNK_NEGOTIATION;
2271 	cmd.code = 0;
2272 	cmd.idx = 0;
2273 	ace_issue_cmd(regs, &cmd);
2274 #endif
2275 
2276 	netif_start_queue(dev);
2277 
2278 	/*
2279 	 * Setup the bottom half rx ring refill handler
2280 	 */
2281 	tasklet_init(&ap->ace_tasklet, ace_tasklet, (unsigned long)dev);
2282 	return 0;
2283 }
2284 
2285 
2286 static int ace_close(struct net_device *dev)
2287 {
2288 	struct ace_private *ap = netdev_priv(dev);
2289 	struct ace_regs __iomem *regs = ap->regs;
2290 	struct cmd cmd;
2291 	unsigned long flags;
2292 	short i;
2293 
2294 	/*
2295 	 * Without (or before) releasing irq and stopping hardware, this
2296 	 * is an absolute non-sense, by the way. It will be reset instantly
2297 	 * by the first irq.
2298 	 */
2299 	netif_stop_queue(dev);
2300 
2301 
2302 	if (ap->promisc) {
2303 		cmd.evt = C_SET_PROMISC_MODE;
2304 		cmd.code = C_C_PROMISC_DISABLE;
2305 		cmd.idx = 0;
2306 		ace_issue_cmd(regs, &cmd);
2307 		ap->promisc = 0;
2308 	}
2309 
2310 	cmd.evt = C_HOST_STATE;
2311 	cmd.code = C_C_STACK_DOWN;
2312 	cmd.idx = 0;
2313 	ace_issue_cmd(regs, &cmd);
2314 
2315 	tasklet_kill(&ap->ace_tasklet);
2316 
2317 	/*
2318 	 * Make sure one CPU is not processing packets while
2319 	 * buffers are being released by another.
2320 	 */
2321 
2322 	local_irq_save(flags);
2323 	ace_mask_irq(dev);
2324 
2325 	for (i = 0; i < ACE_TX_RING_ENTRIES(ap); i++) {
2326 		struct sk_buff *skb;
2327 		struct tx_ring_info *info;
2328 
2329 		info = ap->skb->tx_skbuff + i;
2330 		skb = info->skb;
2331 
2332 		if (dma_unmap_len(info, maplen)) {
2333 			if (ACE_IS_TIGON_I(ap)) {
2334 				/* NB: TIGON_1 is special, tx_ring is in io space */
2335 				struct tx_desc __iomem *tx;
2336 				tx = (__force struct tx_desc __iomem *) &ap->tx_ring[i];
2337 				writel(0, &tx->addr.addrhi);
2338 				writel(0, &tx->addr.addrlo);
2339 				writel(0, &tx->flagsize);
2340 			} else
2341 				memset(ap->tx_ring + i, 0,
2342 				       sizeof(struct tx_desc));
2343 			pci_unmap_page(ap->pdev, dma_unmap_addr(info, mapping),
2344 				       dma_unmap_len(info, maplen),
2345 				       PCI_DMA_TODEVICE);
2346 			dma_unmap_len_set(info, maplen, 0);
2347 		}
2348 		if (skb) {
2349 			dev_kfree_skb(skb);
2350 			info->skb = NULL;
2351 		}
2352 	}
2353 
2354 	if (ap->jumbo) {
2355 		cmd.evt = C_RESET_JUMBO_RNG;
2356 		cmd.code = 0;
2357 		cmd.idx = 0;
2358 		ace_issue_cmd(regs, &cmd);
2359 	}
2360 
2361 	ace_unmask_irq(dev);
2362 	local_irq_restore(flags);
2363 
2364 	return 0;
2365 }
2366 
2367 
2368 static inline dma_addr_t
2369 ace_map_tx_skb(struct ace_private *ap, struct sk_buff *skb,
2370 	       struct sk_buff *tail, u32 idx)
2371 {
2372 	dma_addr_t mapping;
2373 	struct tx_ring_info *info;
2374 
2375 	mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
2376 			       offset_in_page(skb->data),
2377 			       skb->len, PCI_DMA_TODEVICE);
2378 
2379 	info = ap->skb->tx_skbuff + idx;
2380 	info->skb = tail;
2381 	dma_unmap_addr_set(info, mapping, mapping);
2382 	dma_unmap_len_set(info, maplen, skb->len);
2383 	return mapping;
2384 }
2385 
2386 
2387 static inline void
2388 ace_load_tx_bd(struct ace_private *ap, struct tx_desc *desc, u64 addr,
2389 	       u32 flagsize, u32 vlan_tag)
2390 {
2391 #if !USE_TX_COAL_NOW
2392 	flagsize &= ~BD_FLG_COAL_NOW;
2393 #endif
2394 
2395 	if (ACE_IS_TIGON_I(ap)) {
2396 		struct tx_desc __iomem *io = (__force struct tx_desc __iomem *) desc;
2397 		writel(addr >> 32, &io->addr.addrhi);
2398 		writel(addr & 0xffffffff, &io->addr.addrlo);
2399 		writel(flagsize, &io->flagsize);
2400 		writel(vlan_tag, &io->vlanres);
2401 	} else {
2402 		desc->addr.addrhi = addr >> 32;
2403 		desc->addr.addrlo = addr;
2404 		desc->flagsize = flagsize;
2405 		desc->vlanres = vlan_tag;
2406 	}
2407 }
2408 
2409 
2410 static netdev_tx_t ace_start_xmit(struct sk_buff *skb,
2411 				  struct net_device *dev)
2412 {
2413 	struct ace_private *ap = netdev_priv(dev);
2414 	struct ace_regs __iomem *regs = ap->regs;
2415 	struct tx_desc *desc;
2416 	u32 idx, flagsize;
2417 	unsigned long maxjiff = jiffies + 3*HZ;
2418 
2419 restart:
2420 	idx = ap->tx_prd;
2421 
2422 	if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2423 		goto overflow;
2424 
2425 	if (!skb_shinfo(skb)->nr_frags)	{
2426 		dma_addr_t mapping;
2427 		u32 vlan_tag = 0;
2428 
2429 		mapping = ace_map_tx_skb(ap, skb, skb, idx);
2430 		flagsize = (skb->len << 16) | (BD_FLG_END);
2431 		if (skb->ip_summed == CHECKSUM_PARTIAL)
2432 			flagsize |= BD_FLG_TCP_UDP_SUM;
2433 		if (skb_vlan_tag_present(skb)) {
2434 			flagsize |= BD_FLG_VLAN_TAG;
2435 			vlan_tag = skb_vlan_tag_get(skb);
2436 		}
2437 		desc = ap->tx_ring + idx;
2438 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2439 
2440 		/* Look at ace_tx_int for explanations. */
2441 		if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2442 			flagsize |= BD_FLG_COAL_NOW;
2443 
2444 		ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2445 	} else {
2446 		dma_addr_t mapping;
2447 		u32 vlan_tag = 0;
2448 		int i, len = 0;
2449 
2450 		mapping = ace_map_tx_skb(ap, skb, NULL, idx);
2451 		flagsize = (skb_headlen(skb) << 16);
2452 		if (skb->ip_summed == CHECKSUM_PARTIAL)
2453 			flagsize |= BD_FLG_TCP_UDP_SUM;
2454 		if (skb_vlan_tag_present(skb)) {
2455 			flagsize |= BD_FLG_VLAN_TAG;
2456 			vlan_tag = skb_vlan_tag_get(skb);
2457 		}
2458 
2459 		ace_load_tx_bd(ap, ap->tx_ring + idx, mapping, flagsize, vlan_tag);
2460 
2461 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2462 
2463 		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2464 			const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2465 			struct tx_ring_info *info;
2466 
2467 			len += skb_frag_size(frag);
2468 			info = ap->skb->tx_skbuff + idx;
2469 			desc = ap->tx_ring + idx;
2470 
2471 			mapping = skb_frag_dma_map(&ap->pdev->dev, frag, 0,
2472 						   skb_frag_size(frag),
2473 						   DMA_TO_DEVICE);
2474 
2475 			flagsize = skb_frag_size(frag) << 16;
2476 			if (skb->ip_summed == CHECKSUM_PARTIAL)
2477 				flagsize |= BD_FLG_TCP_UDP_SUM;
2478 			idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2479 
2480 			if (i == skb_shinfo(skb)->nr_frags - 1) {
2481 				flagsize |= BD_FLG_END;
2482 				if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2483 					flagsize |= BD_FLG_COAL_NOW;
2484 
2485 				/*
2486 				 * Only the last fragment frees
2487 				 * the skb!
2488 				 */
2489 				info->skb = skb;
2490 			} else {
2491 				info->skb = NULL;
2492 			}
2493 			dma_unmap_addr_set(info, mapping, mapping);
2494 			dma_unmap_len_set(info, maplen, skb_frag_size(frag));
2495 			ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2496 		}
2497 	}
2498 
2499  	wmb();
2500  	ap->tx_prd = idx;
2501  	ace_set_txprd(regs, ap, idx);
2502 
2503 	if (flagsize & BD_FLG_COAL_NOW) {
2504 		netif_stop_queue(dev);
2505 
2506 		/*
2507 		 * A TX-descriptor producer (an IRQ) might have gotten
2508 		 * between, making the ring free again. Since xmit is
2509 		 * serialized, this is the only situation we have to
2510 		 * re-test.
2511 		 */
2512 		if (!tx_ring_full(ap, ap->tx_ret_csm, idx))
2513 			netif_wake_queue(dev);
2514 	}
2515 
2516 	return NETDEV_TX_OK;
2517 
2518 overflow:
2519 	/*
2520 	 * This race condition is unavoidable with lock-free drivers.
2521 	 * We wake up the queue _before_ tx_prd is advanced, so that we can
2522 	 * enter hard_start_xmit too early, while tx ring still looks closed.
2523 	 * This happens ~1-4 times per 100000 packets, so that we can allow
2524 	 * to loop syncing to other CPU. Probably, we need an additional
2525 	 * wmb() in ace_tx_intr as well.
2526 	 *
2527 	 * Note that this race is relieved by reserving one more entry
2528 	 * in tx ring than it is necessary (see original non-SG driver).
2529 	 * However, with SG we need to reserve 2*MAX_SKB_FRAGS+1, which
2530 	 * is already overkill.
2531 	 *
2532 	 * Alternative is to return with 1 not throttling queue. In this
2533 	 * case loop becomes longer, no more useful effects.
2534 	 */
2535 	if (time_before(jiffies, maxjiff)) {
2536 		barrier();
2537 		cpu_relax();
2538 		goto restart;
2539 	}
2540 
2541 	/* The ring is stuck full. */
2542 	printk(KERN_WARNING "%s: Transmit ring stuck full\n", dev->name);
2543 	return NETDEV_TX_BUSY;
2544 }
2545 
2546 
2547 static int ace_change_mtu(struct net_device *dev, int new_mtu)
2548 {
2549 	struct ace_private *ap = netdev_priv(dev);
2550 	struct ace_regs __iomem *regs = ap->regs;
2551 
2552 	writel(new_mtu + ETH_HLEN + 4, &regs->IfMtu);
2553 	dev->mtu = new_mtu;
2554 
2555 	if (new_mtu > ACE_STD_MTU) {
2556 		if (!(ap->jumbo)) {
2557 			printk(KERN_INFO "%s: Enabling Jumbo frame "
2558 			       "support\n", dev->name);
2559 			ap->jumbo = 1;
2560 			if (!test_and_set_bit(0, &ap->jumbo_refill_busy))
2561 				ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE);
2562 			ace_set_rxtx_parms(dev, 1);
2563 		}
2564 	} else {
2565 		while (test_and_set_bit(0, &ap->jumbo_refill_busy));
2566 		ace_sync_irq(dev->irq);
2567 		ace_set_rxtx_parms(dev, 0);
2568 		if (ap->jumbo) {
2569 			struct cmd cmd;
2570 
2571 			cmd.evt = C_RESET_JUMBO_RNG;
2572 			cmd.code = 0;
2573 			cmd.idx = 0;
2574 			ace_issue_cmd(regs, &cmd);
2575 		}
2576 	}
2577 
2578 	return 0;
2579 }
2580 
2581 static int ace_get_link_ksettings(struct net_device *dev,
2582 				  struct ethtool_link_ksettings *cmd)
2583 {
2584 	struct ace_private *ap = netdev_priv(dev);
2585 	struct ace_regs __iomem *regs = ap->regs;
2586 	u32 link;
2587 	u32 supported;
2588 
2589 	memset(cmd, 0, sizeof(struct ethtool_link_ksettings));
2590 
2591 	supported = (SUPPORTED_10baseT_Half | SUPPORTED_10baseT_Full |
2592 		     SUPPORTED_100baseT_Half | SUPPORTED_100baseT_Full |
2593 		     SUPPORTED_1000baseT_Half | SUPPORTED_1000baseT_Full |
2594 		     SUPPORTED_Autoneg | SUPPORTED_FIBRE);
2595 
2596 	cmd->base.port = PORT_FIBRE;
2597 
2598 	link = readl(&regs->GigLnkState);
2599 	if (link & LNK_1000MB) {
2600 		cmd->base.speed = SPEED_1000;
2601 	} else {
2602 		link = readl(&regs->FastLnkState);
2603 		if (link & LNK_100MB)
2604 			cmd->base.speed = SPEED_100;
2605 		else if (link & LNK_10MB)
2606 			cmd->base.speed = SPEED_10;
2607 		else
2608 			cmd->base.speed = 0;
2609 	}
2610 	if (link & LNK_FULL_DUPLEX)
2611 		cmd->base.duplex = DUPLEX_FULL;
2612 	else
2613 		cmd->base.duplex = DUPLEX_HALF;
2614 
2615 	if (link & LNK_NEGOTIATE)
2616 		cmd->base.autoneg = AUTONEG_ENABLE;
2617 	else
2618 		cmd->base.autoneg = AUTONEG_DISABLE;
2619 
2620 #if 0
2621 	/*
2622 	 * Current struct ethtool_cmd is insufficient
2623 	 */
2624 	ecmd->trace = readl(&regs->TuneTrace);
2625 
2626 	ecmd->txcoal = readl(&regs->TuneTxCoalTicks);
2627 	ecmd->rxcoal = readl(&regs->TuneRxCoalTicks);
2628 #endif
2629 
2630 	ethtool_convert_legacy_u32_to_link_mode(cmd->link_modes.supported,
2631 						supported);
2632 
2633 	return 0;
2634 }
2635 
2636 static int ace_set_link_ksettings(struct net_device *dev,
2637 				  const struct ethtool_link_ksettings *cmd)
2638 {
2639 	struct ace_private *ap = netdev_priv(dev);
2640 	struct ace_regs __iomem *regs = ap->regs;
2641 	u32 link, speed;
2642 
2643 	link = readl(&regs->GigLnkState);
2644 	if (link & LNK_1000MB)
2645 		speed = SPEED_1000;
2646 	else {
2647 		link = readl(&regs->FastLnkState);
2648 		if (link & LNK_100MB)
2649 			speed = SPEED_100;
2650 		else if (link & LNK_10MB)
2651 			speed = SPEED_10;
2652 		else
2653 			speed = SPEED_100;
2654 	}
2655 
2656 	link = LNK_ENABLE | LNK_1000MB | LNK_100MB | LNK_10MB |
2657 		LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL;
2658 	if (!ACE_IS_TIGON_I(ap))
2659 		link |= LNK_TX_FLOW_CTL_Y;
2660 	if (cmd->base.autoneg == AUTONEG_ENABLE)
2661 		link |= LNK_NEGOTIATE;
2662 	if (cmd->base.speed != speed) {
2663 		link &= ~(LNK_1000MB | LNK_100MB | LNK_10MB);
2664 		switch (cmd->base.speed) {
2665 		case SPEED_1000:
2666 			link |= LNK_1000MB;
2667 			break;
2668 		case SPEED_100:
2669 			link |= LNK_100MB;
2670 			break;
2671 		case SPEED_10:
2672 			link |= LNK_10MB;
2673 			break;
2674 		}
2675 	}
2676 
2677 	if (cmd->base.duplex == DUPLEX_FULL)
2678 		link |= LNK_FULL_DUPLEX;
2679 
2680 	if (link != ap->link) {
2681 		struct cmd cmd;
2682 		printk(KERN_INFO "%s: Renegotiating link state\n",
2683 		       dev->name);
2684 
2685 		ap->link = link;
2686 		writel(link, &regs->TuneLink);
2687 		if (!ACE_IS_TIGON_I(ap))
2688 			writel(link, &regs->TuneFastLink);
2689 		wmb();
2690 
2691 		cmd.evt = C_LNK_NEGOTIATION;
2692 		cmd.code = 0;
2693 		cmd.idx = 0;
2694 		ace_issue_cmd(regs, &cmd);
2695 	}
2696 	return 0;
2697 }
2698 
2699 static void ace_get_drvinfo(struct net_device *dev,
2700 			    struct ethtool_drvinfo *info)
2701 {
2702 	struct ace_private *ap = netdev_priv(dev);
2703 
2704 	strlcpy(info->driver, "acenic", sizeof(info->driver));
2705 	snprintf(info->version, sizeof(info->version), "%i.%i.%i",
2706 		 ap->firmware_major, ap->firmware_minor,
2707 		 ap->firmware_fix);
2708 
2709 	if (ap->pdev)
2710 		strlcpy(info->bus_info, pci_name(ap->pdev),
2711 			sizeof(info->bus_info));
2712 
2713 }
2714 
2715 /*
2716  * Set the hardware MAC address.
2717  */
2718 static int ace_set_mac_addr(struct net_device *dev, void *p)
2719 {
2720 	struct ace_private *ap = netdev_priv(dev);
2721 	struct ace_regs __iomem *regs = ap->regs;
2722 	struct sockaddr *addr=p;
2723 	u8 *da;
2724 	struct cmd cmd;
2725 
2726 	if(netif_running(dev))
2727 		return -EBUSY;
2728 
2729 	memcpy(dev->dev_addr, addr->sa_data,dev->addr_len);
2730 
2731 	da = (u8 *)dev->dev_addr;
2732 
2733 	writel(da[0] << 8 | da[1], &regs->MacAddrHi);
2734 	writel((da[2] << 24) | (da[3] << 16) | (da[4] << 8) | da[5],
2735 	       &regs->MacAddrLo);
2736 
2737 	cmd.evt = C_SET_MAC_ADDR;
2738 	cmd.code = 0;
2739 	cmd.idx = 0;
2740 	ace_issue_cmd(regs, &cmd);
2741 
2742 	return 0;
2743 }
2744 
2745 
2746 static void ace_set_multicast_list(struct net_device *dev)
2747 {
2748 	struct ace_private *ap = netdev_priv(dev);
2749 	struct ace_regs __iomem *regs = ap->regs;
2750 	struct cmd cmd;
2751 
2752 	if ((dev->flags & IFF_ALLMULTI) && !(ap->mcast_all)) {
2753 		cmd.evt = C_SET_MULTICAST_MODE;
2754 		cmd.code = C_C_MCAST_ENABLE;
2755 		cmd.idx = 0;
2756 		ace_issue_cmd(regs, &cmd);
2757 		ap->mcast_all = 1;
2758 	} else if (ap->mcast_all) {
2759 		cmd.evt = C_SET_MULTICAST_MODE;
2760 		cmd.code = C_C_MCAST_DISABLE;
2761 		cmd.idx = 0;
2762 		ace_issue_cmd(regs, &cmd);
2763 		ap->mcast_all = 0;
2764 	}
2765 
2766 	if ((dev->flags & IFF_PROMISC) && !(ap->promisc)) {
2767 		cmd.evt = C_SET_PROMISC_MODE;
2768 		cmd.code = C_C_PROMISC_ENABLE;
2769 		cmd.idx = 0;
2770 		ace_issue_cmd(regs, &cmd);
2771 		ap->promisc = 1;
2772 	}else if (!(dev->flags & IFF_PROMISC) && (ap->promisc)) {
2773 		cmd.evt = C_SET_PROMISC_MODE;
2774 		cmd.code = C_C_PROMISC_DISABLE;
2775 		cmd.idx = 0;
2776 		ace_issue_cmd(regs, &cmd);
2777 		ap->promisc = 0;
2778 	}
2779 
2780 	/*
2781 	 * For the time being multicast relies on the upper layers
2782 	 * filtering it properly. The Firmware does not allow one to
2783 	 * set the entire multicast list at a time and keeping track of
2784 	 * it here is going to be messy.
2785 	 */
2786 	if (!netdev_mc_empty(dev) && !ap->mcast_all) {
2787 		cmd.evt = C_SET_MULTICAST_MODE;
2788 		cmd.code = C_C_MCAST_ENABLE;
2789 		cmd.idx = 0;
2790 		ace_issue_cmd(regs, &cmd);
2791 	}else if (!ap->mcast_all) {
2792 		cmd.evt = C_SET_MULTICAST_MODE;
2793 		cmd.code = C_C_MCAST_DISABLE;
2794 		cmd.idx = 0;
2795 		ace_issue_cmd(regs, &cmd);
2796 	}
2797 }
2798 
2799 
2800 static struct net_device_stats *ace_get_stats(struct net_device *dev)
2801 {
2802 	struct ace_private *ap = netdev_priv(dev);
2803 	struct ace_mac_stats __iomem *mac_stats =
2804 		(struct ace_mac_stats __iomem *)ap->regs->Stats;
2805 
2806 	dev->stats.rx_missed_errors = readl(&mac_stats->drop_space);
2807 	dev->stats.multicast = readl(&mac_stats->kept_mc);
2808 	dev->stats.collisions = readl(&mac_stats->coll);
2809 
2810 	return &dev->stats;
2811 }
2812 
2813 
2814 static void ace_copy(struct ace_regs __iomem *regs, const __be32 *src,
2815 		     u32 dest, int size)
2816 {
2817 	void __iomem *tdest;
2818 	short tsize, i;
2819 
2820 	if (size <= 0)
2821 		return;
2822 
2823 	while (size > 0) {
2824 		tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2825 			    min_t(u32, size, ACE_WINDOW_SIZE));
2826 		tdest = (void __iomem *) &regs->Window +
2827 			(dest & (ACE_WINDOW_SIZE - 1));
2828 		writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2829 		for (i = 0; i < (tsize / 4); i++) {
2830 			/* Firmware is big-endian */
2831 			writel(be32_to_cpup(src), tdest);
2832 			src++;
2833 			tdest += 4;
2834 			dest += 4;
2835 			size -= 4;
2836 		}
2837 	}
2838 }
2839 
2840 
2841 static void ace_clear(struct ace_regs __iomem *regs, u32 dest, int size)
2842 {
2843 	void __iomem *tdest;
2844 	short tsize = 0, i;
2845 
2846 	if (size <= 0)
2847 		return;
2848 
2849 	while (size > 0) {
2850 		tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2851 				min_t(u32, size, ACE_WINDOW_SIZE));
2852 		tdest = (void __iomem *) &regs->Window +
2853 			(dest & (ACE_WINDOW_SIZE - 1));
2854 		writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2855 
2856 		for (i = 0; i < (tsize / 4); i++) {
2857 			writel(0, tdest + i*4);
2858 		}
2859 
2860 		dest += tsize;
2861 		size -= tsize;
2862 	}
2863 }
2864 
2865 
2866 /*
2867  * Download the firmware into the SRAM on the NIC
2868  *
2869  * This operation requires the NIC to be halted and is performed with
2870  * interrupts disabled and with the spinlock hold.
2871  */
2872 static int ace_load_firmware(struct net_device *dev)
2873 {
2874 	const struct firmware *fw;
2875 	const char *fw_name = "acenic/tg2.bin";
2876 	struct ace_private *ap = netdev_priv(dev);
2877 	struct ace_regs __iomem *regs = ap->regs;
2878 	const __be32 *fw_data;
2879 	u32 load_addr;
2880 	int ret;
2881 
2882 	if (!(readl(&regs->CpuCtrl) & CPU_HALTED)) {
2883 		printk(KERN_ERR "%s: trying to download firmware while the "
2884 		       "CPU is running!\n", ap->name);
2885 		return -EFAULT;
2886 	}
2887 
2888 	if (ACE_IS_TIGON_I(ap))
2889 		fw_name = "acenic/tg1.bin";
2890 
2891 	ret = request_firmware(&fw, fw_name, &ap->pdev->dev);
2892 	if (ret) {
2893 		printk(KERN_ERR "%s: Failed to load firmware \"%s\"\n",
2894 		       ap->name, fw_name);
2895 		return ret;
2896 	}
2897 
2898 	fw_data = (void *)fw->data;
2899 
2900 	/* Firmware blob starts with version numbers, followed by
2901 	   load and start address. Remainder is the blob to be loaded
2902 	   contiguously from load address. We don't bother to represent
2903 	   the BSS/SBSS sections any more, since we were clearing the
2904 	   whole thing anyway. */
2905 	ap->firmware_major = fw->data[0];
2906 	ap->firmware_minor = fw->data[1];
2907 	ap->firmware_fix = fw->data[2];
2908 
2909 	ap->firmware_start = be32_to_cpu(fw_data[1]);
2910 	if (ap->firmware_start < 0x4000 || ap->firmware_start >= 0x80000) {
2911 		printk(KERN_ERR "%s: bogus load address %08x in \"%s\"\n",
2912 		       ap->name, ap->firmware_start, fw_name);
2913 		ret = -EINVAL;
2914 		goto out;
2915 	}
2916 
2917 	load_addr = be32_to_cpu(fw_data[2]);
2918 	if (load_addr < 0x4000 || load_addr >= 0x80000) {
2919 		printk(KERN_ERR "%s: bogus load address %08x in \"%s\"\n",
2920 		       ap->name, load_addr, fw_name);
2921 		ret = -EINVAL;
2922 		goto out;
2923 	}
2924 
2925 	/*
2926 	 * Do not try to clear more than 512KiB or we end up seeing
2927 	 * funny things on NICs with only 512KiB SRAM
2928 	 */
2929 	ace_clear(regs, 0x2000, 0x80000-0x2000);
2930 	ace_copy(regs, &fw_data[3], load_addr, fw->size-12);
2931  out:
2932 	release_firmware(fw);
2933 	return ret;
2934 }
2935 
2936 
2937 /*
2938  * The eeprom on the AceNIC is an Atmel i2c EEPROM.
2939  *
2940  * Accessing the EEPROM is `interesting' to say the least - don't read
2941  * this code right after dinner.
2942  *
2943  * This is all about black magic and bit-banging the device .... I
2944  * wonder in what hospital they have put the guy who designed the i2c
2945  * specs.
2946  *
2947  * Oh yes, this is only the beginning!
2948  *
2949  * Thanks to Stevarino Webinski for helping tracking down the bugs in the
2950  * code i2c readout code by beta testing all my hacks.
2951  */
2952 static void eeprom_start(struct ace_regs __iomem *regs)
2953 {
2954 	u32 local;
2955 
2956 	readl(&regs->LocalCtrl);
2957 	udelay(ACE_SHORT_DELAY);
2958 	local = readl(&regs->LocalCtrl);
2959 	local |= EEPROM_DATA_OUT | EEPROM_WRITE_ENABLE;
2960 	writel(local, &regs->LocalCtrl);
2961 	readl(&regs->LocalCtrl);
2962 	mb();
2963 	udelay(ACE_SHORT_DELAY);
2964 	local |= EEPROM_CLK_OUT;
2965 	writel(local, &regs->LocalCtrl);
2966 	readl(&regs->LocalCtrl);
2967 	mb();
2968 	udelay(ACE_SHORT_DELAY);
2969 	local &= ~EEPROM_DATA_OUT;
2970 	writel(local, &regs->LocalCtrl);
2971 	readl(&regs->LocalCtrl);
2972 	mb();
2973 	udelay(ACE_SHORT_DELAY);
2974 	local &= ~EEPROM_CLK_OUT;
2975 	writel(local, &regs->LocalCtrl);
2976 	readl(&regs->LocalCtrl);
2977 	mb();
2978 }
2979 
2980 
2981 static void eeprom_prep(struct ace_regs __iomem *regs, u8 magic)
2982 {
2983 	short i;
2984 	u32 local;
2985 
2986 	udelay(ACE_SHORT_DELAY);
2987 	local = readl(&regs->LocalCtrl);
2988 	local &= ~EEPROM_DATA_OUT;
2989 	local |= EEPROM_WRITE_ENABLE;
2990 	writel(local, &regs->LocalCtrl);
2991 	readl(&regs->LocalCtrl);
2992 	mb();
2993 
2994 	for (i = 0; i < 8; i++, magic <<= 1) {
2995 		udelay(ACE_SHORT_DELAY);
2996 		if (magic & 0x80)
2997 			local |= EEPROM_DATA_OUT;
2998 		else
2999 			local &= ~EEPROM_DATA_OUT;
3000 		writel(local, &regs->LocalCtrl);
3001 		readl(&regs->LocalCtrl);
3002 		mb();
3003 
3004 		udelay(ACE_SHORT_DELAY);
3005 		local |= EEPROM_CLK_OUT;
3006 		writel(local, &regs->LocalCtrl);
3007 		readl(&regs->LocalCtrl);
3008 		mb();
3009 		udelay(ACE_SHORT_DELAY);
3010 		local &= ~(EEPROM_CLK_OUT | EEPROM_DATA_OUT);
3011 		writel(local, &regs->LocalCtrl);
3012 		readl(&regs->LocalCtrl);
3013 		mb();
3014 	}
3015 }
3016 
3017 
3018 static int eeprom_check_ack(struct ace_regs __iomem *regs)
3019 {
3020 	int state;
3021 	u32 local;
3022 
3023 	local = readl(&regs->LocalCtrl);
3024 	local &= ~EEPROM_WRITE_ENABLE;
3025 	writel(local, &regs->LocalCtrl);
3026 	readl(&regs->LocalCtrl);
3027 	mb();
3028 	udelay(ACE_LONG_DELAY);
3029 	local |= EEPROM_CLK_OUT;
3030 	writel(local, &regs->LocalCtrl);
3031 	readl(&regs->LocalCtrl);
3032 	mb();
3033 	udelay(ACE_SHORT_DELAY);
3034 	/* sample data in middle of high clk */
3035 	state = (readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0;
3036 	udelay(ACE_SHORT_DELAY);
3037 	mb();
3038 	writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3039 	readl(&regs->LocalCtrl);
3040 	mb();
3041 
3042 	return state;
3043 }
3044 
3045 
3046 static void eeprom_stop(struct ace_regs __iomem *regs)
3047 {
3048 	u32 local;
3049 
3050 	udelay(ACE_SHORT_DELAY);
3051 	local = readl(&regs->LocalCtrl);
3052 	local |= EEPROM_WRITE_ENABLE;
3053 	writel(local, &regs->LocalCtrl);
3054 	readl(&regs->LocalCtrl);
3055 	mb();
3056 	udelay(ACE_SHORT_DELAY);
3057 	local &= ~EEPROM_DATA_OUT;
3058 	writel(local, &regs->LocalCtrl);
3059 	readl(&regs->LocalCtrl);
3060 	mb();
3061 	udelay(ACE_SHORT_DELAY);
3062 	local |= EEPROM_CLK_OUT;
3063 	writel(local, &regs->LocalCtrl);
3064 	readl(&regs->LocalCtrl);
3065 	mb();
3066 	udelay(ACE_SHORT_DELAY);
3067 	local |= EEPROM_DATA_OUT;
3068 	writel(local, &regs->LocalCtrl);
3069 	readl(&regs->LocalCtrl);
3070 	mb();
3071 	udelay(ACE_LONG_DELAY);
3072 	local &= ~EEPROM_CLK_OUT;
3073 	writel(local, &regs->LocalCtrl);
3074 	mb();
3075 }
3076 
3077 
3078 /*
3079  * Read a whole byte from the EEPROM.
3080  */
3081 static int read_eeprom_byte(struct net_device *dev, unsigned long offset)
3082 {
3083 	struct ace_private *ap = netdev_priv(dev);
3084 	struct ace_regs __iomem *regs = ap->regs;
3085 	unsigned long flags;
3086 	u32 local;
3087 	int result = 0;
3088 	short i;
3089 
3090 	/*
3091 	 * Don't take interrupts on this CPU will bit banging
3092 	 * the %#%#@$ I2C device
3093 	 */
3094 	local_irq_save(flags);
3095 
3096 	eeprom_start(regs);
3097 
3098 	eeprom_prep(regs, EEPROM_WRITE_SELECT);
3099 	if (eeprom_check_ack(regs)) {
3100 		local_irq_restore(flags);
3101 		printk(KERN_ERR "%s: Unable to sync eeprom\n", ap->name);
3102 		result = -EIO;
3103 		goto eeprom_read_error;
3104 	}
3105 
3106 	eeprom_prep(regs, (offset >> 8) & 0xff);
3107 	if (eeprom_check_ack(regs)) {
3108 		local_irq_restore(flags);
3109 		printk(KERN_ERR "%s: Unable to set address byte 0\n",
3110 		       ap->name);
3111 		result = -EIO;
3112 		goto eeprom_read_error;
3113 	}
3114 
3115 	eeprom_prep(regs, offset & 0xff);
3116 	if (eeprom_check_ack(regs)) {
3117 		local_irq_restore(flags);
3118 		printk(KERN_ERR "%s: Unable to set address byte 1\n",
3119 		       ap->name);
3120 		result = -EIO;
3121 		goto eeprom_read_error;
3122 	}
3123 
3124 	eeprom_start(regs);
3125 	eeprom_prep(regs, EEPROM_READ_SELECT);
3126 	if (eeprom_check_ack(regs)) {
3127 		local_irq_restore(flags);
3128 		printk(KERN_ERR "%s: Unable to set READ_SELECT\n",
3129 		       ap->name);
3130 		result = -EIO;
3131 		goto eeprom_read_error;
3132 	}
3133 
3134 	for (i = 0; i < 8; i++) {
3135 		local = readl(&regs->LocalCtrl);
3136 		local &= ~EEPROM_WRITE_ENABLE;
3137 		writel(local, &regs->LocalCtrl);
3138 		readl(&regs->LocalCtrl);
3139 		udelay(ACE_LONG_DELAY);
3140 		mb();
3141 		local |= EEPROM_CLK_OUT;
3142 		writel(local, &regs->LocalCtrl);
3143 		readl(&regs->LocalCtrl);
3144 		mb();
3145 		udelay(ACE_SHORT_DELAY);
3146 		/* sample data mid high clk */
3147 		result = (result << 1) |
3148 			((readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0);
3149 		udelay(ACE_SHORT_DELAY);
3150 		mb();
3151 		local = readl(&regs->LocalCtrl);
3152 		local &= ~EEPROM_CLK_OUT;
3153 		writel(local, &regs->LocalCtrl);
3154 		readl(&regs->LocalCtrl);
3155 		udelay(ACE_SHORT_DELAY);
3156 		mb();
3157 		if (i == 7) {
3158 			local |= EEPROM_WRITE_ENABLE;
3159 			writel(local, &regs->LocalCtrl);
3160 			readl(&regs->LocalCtrl);
3161 			mb();
3162 			udelay(ACE_SHORT_DELAY);
3163 		}
3164 	}
3165 
3166 	local |= EEPROM_DATA_OUT;
3167 	writel(local, &regs->LocalCtrl);
3168 	readl(&regs->LocalCtrl);
3169 	mb();
3170 	udelay(ACE_SHORT_DELAY);
3171 	writel(readl(&regs->LocalCtrl) | EEPROM_CLK_OUT, &regs->LocalCtrl);
3172 	readl(&regs->LocalCtrl);
3173 	udelay(ACE_LONG_DELAY);
3174 	writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3175 	readl(&regs->LocalCtrl);
3176 	mb();
3177 	udelay(ACE_SHORT_DELAY);
3178 	eeprom_stop(regs);
3179 
3180 	local_irq_restore(flags);
3181  out:
3182 	return result;
3183 
3184  eeprom_read_error:
3185 	printk(KERN_ERR "%s: Unable to read eeprom byte 0x%02lx\n",
3186 	       ap->name, offset);
3187 	goto out;
3188 }
3189 
3190 module_pci_driver(acenic_pci_driver);
3191