1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * acenic.c: Linux driver for the Alteon AceNIC Gigabit Ethernet card
4  *           and other Tigon based cards.
5  *
6  * Copyright 1998-2002 by Jes Sorensen, <jes@trained-monkey.org>.
7  *
8  * Thanks to Alteon and 3Com for providing hardware and documentation
9  * enabling me to write this driver.
10  *
11  * A mailing list for discussing the use of this driver has been
12  * setup, please subscribe to the lists if you have any questions
13  * about the driver. Send mail to linux-acenic-help@sunsite.auc.dk to
14  * see how to subscribe.
15  *
16  * Additional credits:
17  *   Pete Wyckoff <wyckoff@ca.sandia.gov>: Initial Linux/Alpha and trace
18  *       dump support. The trace dump support has not been
19  *       integrated yet however.
20  *   Troy Benjegerdes: Big Endian (PPC) patches.
21  *   Nate Stahl: Better out of memory handling and stats support.
22  *   Aman Singla: Nasty race between interrupt handler and tx code dealing
23  *                with 'testing the tx_ret_csm and setting tx_full'
24  *   David S. Miller <davem@redhat.com>: conversion to new PCI dma mapping
25  *                                       infrastructure and Sparc support
26  *   Pierrick Pinasseau (CERN): For lending me an Ultra 5 to test the
27  *                              driver under Linux/Sparc64
28  *   Matt Domsch <Matt_Domsch@dell.com>: Detect Alteon 1000baseT cards
29  *                                       ETHTOOL_GDRVINFO support
30  *   Chip Salzenberg <chip@valinux.com>: Fix race condition between tx
31  *                                       handler and close() cleanup.
32  *   Ken Aaker <kdaaker@rchland.vnet.ibm.com>: Correct check for whether
33  *                                       memory mapped IO is enabled to
34  *                                       make the driver work on RS/6000.
35  *   Takayoshi Kouchi <kouchi@hpc.bs1.fc.nec.co.jp>: Identifying problem
36  *                                       where the driver would disable
37  *                                       bus master mode if it had to disable
38  *                                       write and invalidate.
39  *   Stephen Hack <stephen_hack@hp.com>: Fixed ace_set_mac_addr for little
40  *                                       endian systems.
41  *   Val Henson <vhenson@esscom.com>:    Reset Jumbo skb producer and
42  *                                       rx producer index when
43  *                                       flushing the Jumbo ring.
44  *   Hans Grobler <grobh@sun.ac.za>:     Memory leak fixes in the
45  *                                       driver init path.
46  *   Grant Grundler <grundler@cup.hp.com>: PCI write posting fixes.
47  */
48 
49 #include <linux/module.h>
50 #include <linux/moduleparam.h>
51 #include <linux/types.h>
52 #include <linux/errno.h>
53 #include <linux/ioport.h>
54 #include <linux/pci.h>
55 #include <linux/dma-mapping.h>
56 #include <linux/kernel.h>
57 #include <linux/netdevice.h>
58 #include <linux/etherdevice.h>
59 #include <linux/skbuff.h>
60 #include <linux/delay.h>
61 #include <linux/mm.h>
62 #include <linux/highmem.h>
63 #include <linux/sockios.h>
64 #include <linux/firmware.h>
65 #include <linux/slab.h>
66 #include <linux/prefetch.h>
67 #include <linux/if_vlan.h>
68 
69 #ifdef SIOCETHTOOL
70 #include <linux/ethtool.h>
71 #endif
72 
73 #include <net/sock.h>
74 #include <net/ip.h>
75 
76 #include <asm/io.h>
77 #include <asm/irq.h>
78 #include <asm/byteorder.h>
79 #include <linux/uaccess.h>
80 
81 
82 #define DRV_NAME "acenic"
83 
84 #undef INDEX_DEBUG
85 
86 #ifdef CONFIG_ACENIC_OMIT_TIGON_I
87 #define ACE_IS_TIGON_I(ap)	0
88 #define ACE_TX_RING_ENTRIES(ap)	MAX_TX_RING_ENTRIES
89 #else
90 #define ACE_IS_TIGON_I(ap)	(ap->version == 1)
91 #define ACE_TX_RING_ENTRIES(ap)	ap->tx_ring_entries
92 #endif
93 
94 #ifndef PCI_VENDOR_ID_ALTEON
95 #define PCI_VENDOR_ID_ALTEON		0x12ae
96 #endif
97 #ifndef PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE
98 #define PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE  0x0001
99 #define PCI_DEVICE_ID_ALTEON_ACENIC_COPPER 0x0002
100 #endif
101 #ifndef PCI_DEVICE_ID_3COM_3C985
102 #define PCI_DEVICE_ID_3COM_3C985	0x0001
103 #endif
104 #ifndef PCI_VENDOR_ID_NETGEAR
105 #define PCI_VENDOR_ID_NETGEAR		0x1385
106 #define PCI_DEVICE_ID_NETGEAR_GA620	0x620a
107 #endif
108 #ifndef PCI_DEVICE_ID_NETGEAR_GA620T
109 #define PCI_DEVICE_ID_NETGEAR_GA620T	0x630a
110 #endif
111 
112 
113 /*
114  * Farallon used the DEC vendor ID by mistake and they seem not
115  * to care - stinky!
116  */
117 #ifndef PCI_DEVICE_ID_FARALLON_PN9000SX
118 #define PCI_DEVICE_ID_FARALLON_PN9000SX	0x1a
119 #endif
120 #ifndef PCI_DEVICE_ID_FARALLON_PN9100T
121 #define PCI_DEVICE_ID_FARALLON_PN9100T  0xfa
122 #endif
123 #ifndef PCI_VENDOR_ID_SGI
124 #define PCI_VENDOR_ID_SGI		0x10a9
125 #endif
126 #ifndef PCI_DEVICE_ID_SGI_ACENIC
127 #define PCI_DEVICE_ID_SGI_ACENIC	0x0009
128 #endif
129 
130 static const struct pci_device_id acenic_pci_tbl[] = {
131 	{ PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE,
132 	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
133 	{ PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_ALTEON_ACENIC_COPPER,
134 	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
135 	{ PCI_VENDOR_ID_3COM, PCI_DEVICE_ID_3COM_3C985,
136 	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
137 	{ PCI_VENDOR_ID_NETGEAR, PCI_DEVICE_ID_NETGEAR_GA620,
138 	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
139 	{ PCI_VENDOR_ID_NETGEAR, PCI_DEVICE_ID_NETGEAR_GA620T,
140 	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
141 	/*
142 	 * Farallon used the DEC vendor ID on their cards incorrectly,
143 	 * then later Alteon's ID.
144 	 */
145 	{ PCI_VENDOR_ID_DEC, PCI_DEVICE_ID_FARALLON_PN9000SX,
146 	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
147 	{ PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_FARALLON_PN9100T,
148 	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
149 	{ PCI_VENDOR_ID_SGI, PCI_DEVICE_ID_SGI_ACENIC,
150 	  PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
151 	{ }
152 };
153 MODULE_DEVICE_TABLE(pci, acenic_pci_tbl);
154 
155 #define ace_sync_irq(irq)	synchronize_irq(irq)
156 
157 #ifndef offset_in_page
158 #define offset_in_page(ptr)	((unsigned long)(ptr) & ~PAGE_MASK)
159 #endif
160 
161 #define ACE_MAX_MOD_PARMS	8
162 #define BOARD_IDX_STATIC	0
163 #define BOARD_IDX_OVERFLOW	-1
164 
165 #include "acenic.h"
166 
167 /*
168  * These must be defined before the firmware is included.
169  */
170 #define MAX_TEXT_LEN	96*1024
171 #define MAX_RODATA_LEN	8*1024
172 #define MAX_DATA_LEN	2*1024
173 
174 #ifndef tigon2FwReleaseLocal
175 #define tigon2FwReleaseLocal 0
176 #endif
177 
178 /*
179  * This driver currently supports Tigon I and Tigon II based cards
180  * including the Alteon AceNIC, the 3Com 3C985[B] and NetGear
181  * GA620. The driver should also work on the SGI, DEC and Farallon
182  * versions of the card, however I have not been able to test that
183  * myself.
184  *
185  * This card is really neat, it supports receive hardware checksumming
186  * and jumbo frames (up to 9000 bytes) and does a lot of work in the
187  * firmware. Also the programming interface is quite neat, except for
188  * the parts dealing with the i2c eeprom on the card ;-)
189  *
190  * Using jumbo frames:
191  *
192  * To enable jumbo frames, simply specify an mtu between 1500 and 9000
193  * bytes to ifconfig. Jumbo frames can be enabled or disabled at any time
194  * by running `ifconfig eth<X> mtu <MTU>' with <X> being the Ethernet
195  * interface number and <MTU> being the MTU value.
196  *
197  * Module parameters:
198  *
199  * When compiled as a loadable module, the driver allows for a number
200  * of module parameters to be specified. The driver supports the
201  * following module parameters:
202  *
203  *  trace=<val> - Firmware trace level. This requires special traced
204  *                firmware to replace the firmware supplied with
205  *                the driver - for debugging purposes only.
206  *
207  *  link=<val>  - Link state. Normally you want to use the default link
208  *                parameters set by the driver. This can be used to
209  *                override these in case your switch doesn't negotiate
210  *                the link properly. Valid values are:
211  *         0x0001 - Force half duplex link.
212  *         0x0002 - Do not negotiate line speed with the other end.
213  *         0x0010 - 10Mbit/sec link.
214  *         0x0020 - 100Mbit/sec link.
215  *         0x0040 - 1000Mbit/sec link.
216  *         0x0100 - Do not negotiate flow control.
217  *         0x0200 - Enable RX flow control Y
218  *         0x0400 - Enable TX flow control Y (Tigon II NICs only).
219  *                Default value is 0x0270, ie. enable link+flow
220  *                control negotiation. Negotiating the highest
221  *                possible link speed with RX flow control enabled.
222  *
223  *                When disabling link speed negotiation, only one link
224  *                speed is allowed to be specified!
225  *
226  *  tx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
227  *                to wait for more packets to arive before
228  *                interrupting the host, from the time the first
229  *                packet arrives.
230  *
231  *  rx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
232  *                to wait for more packets to arive in the transmit ring,
233  *                before interrupting the host, after transmitting the
234  *                first packet in the ring.
235  *
236  *  max_tx_desc=<val> - maximum number of transmit descriptors
237  *                (packets) transmitted before interrupting the host.
238  *
239  *  max_rx_desc=<val> - maximum number of receive descriptors
240  *                (packets) received before interrupting the host.
241  *
242  *  tx_ratio=<val> - 7 bit value (0 - 63) specifying the split in 64th
243  *                increments of the NIC's on board memory to be used for
244  *                transmit and receive buffers. For the 1MB NIC app. 800KB
245  *                is available, on the 1/2MB NIC app. 300KB is available.
246  *                68KB will always be available as a minimum for both
247  *                directions. The default value is a 50/50 split.
248  *  dis_pci_mem_inval=<val> - disable PCI memory write and invalidate
249  *                operations, default (1) is to always disable this as
250  *                that is what Alteon does on NT. I have not been able
251  *                to measure any real performance differences with
252  *                this on my systems. Set <val>=0 if you want to
253  *                enable these operations.
254  *
255  * If you use more than one NIC, specify the parameters for the
256  * individual NICs with a comma, ie. trace=0,0x00001fff,0 you want to
257  * run tracing on NIC #2 but not on NIC #1 and #3.
258  *
259  * TODO:
260  *
261  * - Proper multicast support.
262  * - NIC dump support.
263  * - More tuning parameters.
264  *
265  * The mini ring is not used under Linux and I am not sure it makes sense
266  * to actually use it.
267  *
268  * New interrupt handler strategy:
269  *
270  * The old interrupt handler worked using the traditional method of
271  * replacing an skbuff with a new one when a packet arrives. However
272  * the rx rings do not need to contain a static number of buffer
273  * descriptors, thus it makes sense to move the memory allocation out
274  * of the main interrupt handler and do it in a bottom half handler
275  * and only allocate new buffers when the number of buffers in the
276  * ring is below a certain threshold. In order to avoid starving the
277  * NIC under heavy load it is however necessary to force allocation
278  * when hitting a minimum threshold. The strategy for alloction is as
279  * follows:
280  *
281  *     RX_LOW_BUF_THRES    - allocate buffers in the bottom half
282  *     RX_PANIC_LOW_THRES  - we are very low on buffers, allocate
283  *                           the buffers in the interrupt handler
284  *     RX_RING_THRES       - maximum number of buffers in the rx ring
285  *     RX_MINI_THRES       - maximum number of buffers in the mini ring
286  *     RX_JUMBO_THRES      - maximum number of buffers in the jumbo ring
287  *
288  * One advantagous side effect of this allocation approach is that the
289  * entire rx processing can be done without holding any spin lock
290  * since the rx rings and registers are totally independent of the tx
291  * ring and its registers.  This of course includes the kmalloc's of
292  * new skb's. Thus start_xmit can run in parallel with rx processing
293  * and the memory allocation on SMP systems.
294  *
295  * Note that running the skb reallocation in a bottom half opens up
296  * another can of races which needs to be handled properly. In
297  * particular it can happen that the interrupt handler tries to run
298  * the reallocation while the bottom half is either running on another
299  * CPU or was interrupted on the same CPU. To get around this the
300  * driver uses bitops to prevent the reallocation routines from being
301  * reentered.
302  *
303  * TX handling can also be done without holding any spin lock, wheee
304  * this is fun! since tx_ret_csm is only written to by the interrupt
305  * handler. The case to be aware of is when shutting down the device
306  * and cleaning up where it is necessary to make sure that
307  * start_xmit() is not running while this is happening. Well DaveM
308  * informs me that this case is already protected against ... bye bye
309  * Mr. Spin Lock, it was nice to know you.
310  *
311  * TX interrupts are now partly disabled so the NIC will only generate
312  * TX interrupts for the number of coal ticks, not for the number of
313  * TX packets in the queue. This should reduce the number of TX only,
314  * ie. when no RX processing is done, interrupts seen.
315  */
316 
317 /*
318  * Threshold values for RX buffer allocation - the low water marks for
319  * when to start refilling the rings are set to 75% of the ring
320  * sizes. It seems to make sense to refill the rings entirely from the
321  * intrrupt handler once it gets below the panic threshold, that way
322  * we don't risk that the refilling is moved to another CPU when the
323  * one running the interrupt handler just got the slab code hot in its
324  * cache.
325  */
326 #define RX_RING_SIZE		72
327 #define RX_MINI_SIZE		64
328 #define RX_JUMBO_SIZE		48
329 
330 #define RX_PANIC_STD_THRES	16
331 #define RX_PANIC_STD_REFILL	(3*RX_PANIC_STD_THRES)/2
332 #define RX_LOW_STD_THRES	(3*RX_RING_SIZE)/4
333 #define RX_PANIC_MINI_THRES	12
334 #define RX_PANIC_MINI_REFILL	(3*RX_PANIC_MINI_THRES)/2
335 #define RX_LOW_MINI_THRES	(3*RX_MINI_SIZE)/4
336 #define RX_PANIC_JUMBO_THRES	6
337 #define RX_PANIC_JUMBO_REFILL	(3*RX_PANIC_JUMBO_THRES)/2
338 #define RX_LOW_JUMBO_THRES	(3*RX_JUMBO_SIZE)/4
339 
340 
341 /*
342  * Size of the mini ring entries, basically these just should be big
343  * enough to take TCP ACKs
344  */
345 #define ACE_MINI_SIZE		100
346 
347 #define ACE_MINI_BUFSIZE	ACE_MINI_SIZE
348 #define ACE_STD_BUFSIZE		(ACE_STD_MTU + ETH_HLEN + 4)
349 #define ACE_JUMBO_BUFSIZE	(ACE_JUMBO_MTU + ETH_HLEN + 4)
350 
351 /*
352  * There seems to be a magic difference in the effect between 995 and 996
353  * but little difference between 900 and 995 ... no idea why.
354  *
355  * There is now a default set of tuning parameters which is set, depending
356  * on whether or not the user enables Jumbo frames. It's assumed that if
357  * Jumbo frames are enabled, the user wants optimal tuning for that case.
358  */
359 #define DEF_TX_COAL		400 /* 996 */
360 #define DEF_TX_MAX_DESC		60  /* was 40 */
361 #define DEF_RX_COAL		120 /* 1000 */
362 #define DEF_RX_MAX_DESC		25
363 #define DEF_TX_RATIO		21 /* 24 */
364 
365 #define DEF_JUMBO_TX_COAL	20
366 #define DEF_JUMBO_TX_MAX_DESC	60
367 #define DEF_JUMBO_RX_COAL	30
368 #define DEF_JUMBO_RX_MAX_DESC	6
369 #define DEF_JUMBO_TX_RATIO	21
370 
371 #if tigon2FwReleaseLocal < 20001118
372 /*
373  * Standard firmware and early modifications duplicate
374  * IRQ load without this flag (coal timer is never reset).
375  * Note that with this flag tx_coal should be less than
376  * time to xmit full tx ring.
377  * 400usec is not so bad for tx ring size of 128.
378  */
379 #define TX_COAL_INTS_ONLY	1	/* worth it */
380 #else
381 /*
382  * With modified firmware, this is not necessary, but still useful.
383  */
384 #define TX_COAL_INTS_ONLY	1
385 #endif
386 
387 #define DEF_TRACE		0
388 #define DEF_STAT		(2 * TICKS_PER_SEC)
389 
390 
391 static int link_state[ACE_MAX_MOD_PARMS];
392 static int trace[ACE_MAX_MOD_PARMS];
393 static int tx_coal_tick[ACE_MAX_MOD_PARMS];
394 static int rx_coal_tick[ACE_MAX_MOD_PARMS];
395 static int max_tx_desc[ACE_MAX_MOD_PARMS];
396 static int max_rx_desc[ACE_MAX_MOD_PARMS];
397 static int tx_ratio[ACE_MAX_MOD_PARMS];
398 static int dis_pci_mem_inval[ACE_MAX_MOD_PARMS] = {1, 1, 1, 1, 1, 1, 1, 1};
399 
400 MODULE_AUTHOR("Jes Sorensen <jes@trained-monkey.org>");
401 MODULE_LICENSE("GPL");
402 MODULE_DESCRIPTION("AceNIC/3C985/GA620 Gigabit Ethernet driver");
403 #ifndef CONFIG_ACENIC_OMIT_TIGON_I
404 MODULE_FIRMWARE("acenic/tg1.bin");
405 #endif
406 MODULE_FIRMWARE("acenic/tg2.bin");
407 
408 module_param_array_named(link, link_state, int, NULL, 0);
409 module_param_array(trace, int, NULL, 0);
410 module_param_array(tx_coal_tick, int, NULL, 0);
411 module_param_array(max_tx_desc, int, NULL, 0);
412 module_param_array(rx_coal_tick, int, NULL, 0);
413 module_param_array(max_rx_desc, int, NULL, 0);
414 module_param_array(tx_ratio, int, NULL, 0);
415 MODULE_PARM_DESC(link, "AceNIC/3C985/NetGear link state");
416 MODULE_PARM_DESC(trace, "AceNIC/3C985/NetGear firmware trace level");
417 MODULE_PARM_DESC(tx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first tx descriptor arrives");
418 MODULE_PARM_DESC(max_tx_desc, "AceNIC/3C985/GA620 max number of transmit descriptors to wait");
419 MODULE_PARM_DESC(rx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first rx descriptor arrives");
420 MODULE_PARM_DESC(max_rx_desc, "AceNIC/3C985/GA620 max number of receive descriptors to wait");
421 MODULE_PARM_DESC(tx_ratio, "AceNIC/3C985/GA620 ratio of NIC memory used for TX/RX descriptors (range 0-63)");
422 
423 
424 static const char version[] =
425   "acenic.c: v0.92 08/05/2002  Jes Sorensen, linux-acenic@SunSITE.dk\n"
426   "                            http://home.cern.ch/~jes/gige/acenic.html\n";
427 
428 static int ace_get_link_ksettings(struct net_device *,
429 				  struct ethtool_link_ksettings *);
430 static int ace_set_link_ksettings(struct net_device *,
431 				  const struct ethtool_link_ksettings *);
432 static void ace_get_drvinfo(struct net_device *, struct ethtool_drvinfo *);
433 
434 static const struct ethtool_ops ace_ethtool_ops = {
435 	.get_drvinfo = ace_get_drvinfo,
436 	.get_link_ksettings = ace_get_link_ksettings,
437 	.set_link_ksettings = ace_set_link_ksettings,
438 };
439 
440 static void ace_watchdog(struct net_device *dev, unsigned int txqueue);
441 
442 static const struct net_device_ops ace_netdev_ops = {
443 	.ndo_open		= ace_open,
444 	.ndo_stop		= ace_close,
445 	.ndo_tx_timeout		= ace_watchdog,
446 	.ndo_get_stats		= ace_get_stats,
447 	.ndo_start_xmit		= ace_start_xmit,
448 	.ndo_set_rx_mode	= ace_set_multicast_list,
449 	.ndo_validate_addr	= eth_validate_addr,
450 	.ndo_set_mac_address	= ace_set_mac_addr,
451 	.ndo_change_mtu		= ace_change_mtu,
452 };
453 
454 static int acenic_probe_one(struct pci_dev *pdev,
455 			    const struct pci_device_id *id)
456 {
457 	struct net_device *dev;
458 	struct ace_private *ap;
459 	static int boards_found;
460 
461 	dev = alloc_etherdev(sizeof(struct ace_private));
462 	if (dev == NULL)
463 		return -ENOMEM;
464 
465 	SET_NETDEV_DEV(dev, &pdev->dev);
466 
467 	ap = netdev_priv(dev);
468 	ap->pdev = pdev;
469 	ap->name = pci_name(pdev);
470 
471 	dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
472 	dev->features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX;
473 
474 	dev->watchdog_timeo = 5*HZ;
475 	dev->min_mtu = 0;
476 	dev->max_mtu = ACE_JUMBO_MTU;
477 
478 	dev->netdev_ops = &ace_netdev_ops;
479 	dev->ethtool_ops = &ace_ethtool_ops;
480 
481 	/* we only display this string ONCE */
482 	if (!boards_found)
483 		printk(version);
484 
485 	if (pci_enable_device(pdev))
486 		goto fail_free_netdev;
487 
488 	/*
489 	 * Enable master mode before we start playing with the
490 	 * pci_command word since pci_set_master() will modify
491 	 * it.
492 	 */
493 	pci_set_master(pdev);
494 
495 	pci_read_config_word(pdev, PCI_COMMAND, &ap->pci_command);
496 
497 	/* OpenFirmware on Mac's does not set this - DOH.. */
498 	if (!(ap->pci_command & PCI_COMMAND_MEMORY)) {
499 		printk(KERN_INFO "%s: Enabling PCI Memory Mapped "
500 		       "access - was not enabled by BIOS/Firmware\n",
501 		       ap->name);
502 		ap->pci_command = ap->pci_command | PCI_COMMAND_MEMORY;
503 		pci_write_config_word(ap->pdev, PCI_COMMAND,
504 				      ap->pci_command);
505 		wmb();
506 	}
507 
508 	pci_read_config_byte(pdev, PCI_LATENCY_TIMER, &ap->pci_latency);
509 	if (ap->pci_latency <= 0x40) {
510 		ap->pci_latency = 0x40;
511 		pci_write_config_byte(pdev, PCI_LATENCY_TIMER, ap->pci_latency);
512 	}
513 
514 	/*
515 	 * Remap the regs into kernel space - this is abuse of
516 	 * dev->base_addr since it was means for I/O port
517 	 * addresses but who gives a damn.
518 	 */
519 	dev->base_addr = pci_resource_start(pdev, 0);
520 	ap->regs = ioremap(dev->base_addr, 0x4000);
521 	if (!ap->regs) {
522 		printk(KERN_ERR "%s:  Unable to map I/O register, "
523 		       "AceNIC %i will be disabled.\n",
524 		       ap->name, boards_found);
525 		goto fail_free_netdev;
526 	}
527 
528 	switch(pdev->vendor) {
529 	case PCI_VENDOR_ID_ALTEON:
530 		if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9100T) {
531 			printk(KERN_INFO "%s: Farallon PN9100-T ",
532 			       ap->name);
533 		} else {
534 			printk(KERN_INFO "%s: Alteon AceNIC ",
535 			       ap->name);
536 		}
537 		break;
538 	case PCI_VENDOR_ID_3COM:
539 		printk(KERN_INFO "%s: 3Com 3C985 ", ap->name);
540 		break;
541 	case PCI_VENDOR_ID_NETGEAR:
542 		printk(KERN_INFO "%s: NetGear GA620 ", ap->name);
543 		break;
544 	case PCI_VENDOR_ID_DEC:
545 		if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9000SX) {
546 			printk(KERN_INFO "%s: Farallon PN9000-SX ",
547 			       ap->name);
548 			break;
549 		}
550 		fallthrough;
551 	case PCI_VENDOR_ID_SGI:
552 		printk(KERN_INFO "%s: SGI AceNIC ", ap->name);
553 		break;
554 	default:
555 		printk(KERN_INFO "%s: Unknown AceNIC ", ap->name);
556 		break;
557 	}
558 
559 	printk("Gigabit Ethernet at 0x%08lx, ", dev->base_addr);
560 	printk("irq %d\n", pdev->irq);
561 
562 #ifdef CONFIG_ACENIC_OMIT_TIGON_I
563 	if ((readl(&ap->regs->HostCtrl) >> 28) == 4) {
564 		printk(KERN_ERR "%s: Driver compiled without Tigon I"
565 		       " support - NIC disabled\n", dev->name);
566 		goto fail_uninit;
567 	}
568 #endif
569 
570 	if (ace_allocate_descriptors(dev))
571 		goto fail_free_netdev;
572 
573 #ifdef MODULE
574 	if (boards_found >= ACE_MAX_MOD_PARMS)
575 		ap->board_idx = BOARD_IDX_OVERFLOW;
576 	else
577 		ap->board_idx = boards_found;
578 #else
579 	ap->board_idx = BOARD_IDX_STATIC;
580 #endif
581 
582 	if (ace_init(dev))
583 		goto fail_free_netdev;
584 
585 	if (register_netdev(dev)) {
586 		printk(KERN_ERR "acenic: device registration failed\n");
587 		goto fail_uninit;
588 	}
589 	ap->name = dev->name;
590 
591 	if (ap->pci_using_dac)
592 		dev->features |= NETIF_F_HIGHDMA;
593 
594 	pci_set_drvdata(pdev, dev);
595 
596 	boards_found++;
597 	return 0;
598 
599  fail_uninit:
600 	ace_init_cleanup(dev);
601  fail_free_netdev:
602 	free_netdev(dev);
603 	return -ENODEV;
604 }
605 
606 static void acenic_remove_one(struct pci_dev *pdev)
607 {
608 	struct net_device *dev = pci_get_drvdata(pdev);
609 	struct ace_private *ap = netdev_priv(dev);
610 	struct ace_regs __iomem *regs = ap->regs;
611 	short i;
612 
613 	unregister_netdev(dev);
614 
615 	writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
616 	if (ap->version >= 2)
617 		writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
618 
619 	/*
620 	 * This clears any pending interrupts
621 	 */
622 	writel(1, &regs->Mb0Lo);
623 	readl(&regs->CpuCtrl);	/* flush */
624 
625 	/*
626 	 * Make sure no other CPUs are processing interrupts
627 	 * on the card before the buffers are being released.
628 	 * Otherwise one might experience some `interesting'
629 	 * effects.
630 	 *
631 	 * Then release the RX buffers - jumbo buffers were
632 	 * already released in ace_close().
633 	 */
634 	ace_sync_irq(dev->irq);
635 
636 	for (i = 0; i < RX_STD_RING_ENTRIES; i++) {
637 		struct sk_buff *skb = ap->skb->rx_std_skbuff[i].skb;
638 
639 		if (skb) {
640 			struct ring_info *ringp;
641 			dma_addr_t mapping;
642 
643 			ringp = &ap->skb->rx_std_skbuff[i];
644 			mapping = dma_unmap_addr(ringp, mapping);
645 			dma_unmap_page(&ap->pdev->dev, mapping,
646 				       ACE_STD_BUFSIZE, DMA_FROM_DEVICE);
647 
648 			ap->rx_std_ring[i].size = 0;
649 			ap->skb->rx_std_skbuff[i].skb = NULL;
650 			dev_kfree_skb(skb);
651 		}
652 	}
653 
654 	if (ap->version >= 2) {
655 		for (i = 0; i < RX_MINI_RING_ENTRIES; i++) {
656 			struct sk_buff *skb = ap->skb->rx_mini_skbuff[i].skb;
657 
658 			if (skb) {
659 				struct ring_info *ringp;
660 				dma_addr_t mapping;
661 
662 				ringp = &ap->skb->rx_mini_skbuff[i];
663 				mapping = dma_unmap_addr(ringp,mapping);
664 				dma_unmap_page(&ap->pdev->dev, mapping,
665 					       ACE_MINI_BUFSIZE,
666 					       DMA_FROM_DEVICE);
667 
668 				ap->rx_mini_ring[i].size = 0;
669 				ap->skb->rx_mini_skbuff[i].skb = NULL;
670 				dev_kfree_skb(skb);
671 			}
672 		}
673 	}
674 
675 	for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
676 		struct sk_buff *skb = ap->skb->rx_jumbo_skbuff[i].skb;
677 		if (skb) {
678 			struct ring_info *ringp;
679 			dma_addr_t mapping;
680 
681 			ringp = &ap->skb->rx_jumbo_skbuff[i];
682 			mapping = dma_unmap_addr(ringp, mapping);
683 			dma_unmap_page(&ap->pdev->dev, mapping,
684 				       ACE_JUMBO_BUFSIZE, DMA_FROM_DEVICE);
685 
686 			ap->rx_jumbo_ring[i].size = 0;
687 			ap->skb->rx_jumbo_skbuff[i].skb = NULL;
688 			dev_kfree_skb(skb);
689 		}
690 	}
691 
692 	ace_init_cleanup(dev);
693 	free_netdev(dev);
694 }
695 
696 static struct pci_driver acenic_pci_driver = {
697 	.name		= "acenic",
698 	.id_table	= acenic_pci_tbl,
699 	.probe		= acenic_probe_one,
700 	.remove		= acenic_remove_one,
701 };
702 
703 static void ace_free_descriptors(struct net_device *dev)
704 {
705 	struct ace_private *ap = netdev_priv(dev);
706 	int size;
707 
708 	if (ap->rx_std_ring != NULL) {
709 		size = (sizeof(struct rx_desc) *
710 			(RX_STD_RING_ENTRIES +
711 			 RX_JUMBO_RING_ENTRIES +
712 			 RX_MINI_RING_ENTRIES +
713 			 RX_RETURN_RING_ENTRIES));
714 		dma_free_coherent(&ap->pdev->dev, size, ap->rx_std_ring,
715 				  ap->rx_ring_base_dma);
716 		ap->rx_std_ring = NULL;
717 		ap->rx_jumbo_ring = NULL;
718 		ap->rx_mini_ring = NULL;
719 		ap->rx_return_ring = NULL;
720 	}
721 	if (ap->evt_ring != NULL) {
722 		size = (sizeof(struct event) * EVT_RING_ENTRIES);
723 		dma_free_coherent(&ap->pdev->dev, size, ap->evt_ring,
724 				  ap->evt_ring_dma);
725 		ap->evt_ring = NULL;
726 	}
727 	if (ap->tx_ring != NULL && !ACE_IS_TIGON_I(ap)) {
728 		size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
729 		dma_free_coherent(&ap->pdev->dev, size, ap->tx_ring,
730 				  ap->tx_ring_dma);
731 	}
732 	ap->tx_ring = NULL;
733 
734 	if (ap->evt_prd != NULL) {
735 		dma_free_coherent(&ap->pdev->dev, sizeof(u32),
736 				  (void *)ap->evt_prd, ap->evt_prd_dma);
737 		ap->evt_prd = NULL;
738 	}
739 	if (ap->rx_ret_prd != NULL) {
740 		dma_free_coherent(&ap->pdev->dev, sizeof(u32),
741 				  (void *)ap->rx_ret_prd, ap->rx_ret_prd_dma);
742 		ap->rx_ret_prd = NULL;
743 	}
744 	if (ap->tx_csm != NULL) {
745 		dma_free_coherent(&ap->pdev->dev, sizeof(u32),
746 				  (void *)ap->tx_csm, ap->tx_csm_dma);
747 		ap->tx_csm = NULL;
748 	}
749 }
750 
751 
752 static int ace_allocate_descriptors(struct net_device *dev)
753 {
754 	struct ace_private *ap = netdev_priv(dev);
755 	int size;
756 
757 	size = (sizeof(struct rx_desc) *
758 		(RX_STD_RING_ENTRIES +
759 		 RX_JUMBO_RING_ENTRIES +
760 		 RX_MINI_RING_ENTRIES +
761 		 RX_RETURN_RING_ENTRIES));
762 
763 	ap->rx_std_ring = dma_alloc_coherent(&ap->pdev->dev, size,
764 					     &ap->rx_ring_base_dma, GFP_KERNEL);
765 	if (ap->rx_std_ring == NULL)
766 		goto fail;
767 
768 	ap->rx_jumbo_ring = ap->rx_std_ring + RX_STD_RING_ENTRIES;
769 	ap->rx_mini_ring = ap->rx_jumbo_ring + RX_JUMBO_RING_ENTRIES;
770 	ap->rx_return_ring = ap->rx_mini_ring + RX_MINI_RING_ENTRIES;
771 
772 	size = (sizeof(struct event) * EVT_RING_ENTRIES);
773 
774 	ap->evt_ring = dma_alloc_coherent(&ap->pdev->dev, size,
775 					  &ap->evt_ring_dma, GFP_KERNEL);
776 
777 	if (ap->evt_ring == NULL)
778 		goto fail;
779 
780 	/*
781 	 * Only allocate a host TX ring for the Tigon II, the Tigon I
782 	 * has to use PCI registers for this ;-(
783 	 */
784 	if (!ACE_IS_TIGON_I(ap)) {
785 		size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
786 
787 		ap->tx_ring = dma_alloc_coherent(&ap->pdev->dev, size,
788 						 &ap->tx_ring_dma, GFP_KERNEL);
789 
790 		if (ap->tx_ring == NULL)
791 			goto fail;
792 	}
793 
794 	ap->evt_prd = dma_alloc_coherent(&ap->pdev->dev, sizeof(u32),
795 					 &ap->evt_prd_dma, GFP_KERNEL);
796 	if (ap->evt_prd == NULL)
797 		goto fail;
798 
799 	ap->rx_ret_prd = dma_alloc_coherent(&ap->pdev->dev, sizeof(u32),
800 					    &ap->rx_ret_prd_dma, GFP_KERNEL);
801 	if (ap->rx_ret_prd == NULL)
802 		goto fail;
803 
804 	ap->tx_csm = dma_alloc_coherent(&ap->pdev->dev, sizeof(u32),
805 					&ap->tx_csm_dma, GFP_KERNEL);
806 	if (ap->tx_csm == NULL)
807 		goto fail;
808 
809 	return 0;
810 
811 fail:
812 	/* Clean up. */
813 	ace_init_cleanup(dev);
814 	return 1;
815 }
816 
817 
818 /*
819  * Generic cleanup handling data allocated during init. Used when the
820  * module is unloaded or if an error occurs during initialization
821  */
822 static void ace_init_cleanup(struct net_device *dev)
823 {
824 	struct ace_private *ap;
825 
826 	ap = netdev_priv(dev);
827 
828 	ace_free_descriptors(dev);
829 
830 	if (ap->info)
831 		dma_free_coherent(&ap->pdev->dev, sizeof(struct ace_info),
832 				  ap->info, ap->info_dma);
833 	kfree(ap->skb);
834 	kfree(ap->trace_buf);
835 
836 	if (dev->irq)
837 		free_irq(dev->irq, dev);
838 
839 	iounmap(ap->regs);
840 }
841 
842 
843 /*
844  * Commands are considered to be slow.
845  */
846 static inline void ace_issue_cmd(struct ace_regs __iomem *regs, struct cmd *cmd)
847 {
848 	u32 idx;
849 
850 	idx = readl(&regs->CmdPrd);
851 
852 	writel(*(u32 *)(cmd), &regs->CmdRng[idx]);
853 	idx = (idx + 1) % CMD_RING_ENTRIES;
854 
855 	writel(idx, &regs->CmdPrd);
856 }
857 
858 
859 static int ace_init(struct net_device *dev)
860 {
861 	struct ace_private *ap;
862 	struct ace_regs __iomem *regs;
863 	struct ace_info *info = NULL;
864 	struct pci_dev *pdev;
865 	unsigned long myjif;
866 	u64 tmp_ptr;
867 	u32 tig_ver, mac1, mac2, tmp, pci_state;
868 	int board_idx, ecode = 0;
869 	short i;
870 	unsigned char cache_size;
871 
872 	ap = netdev_priv(dev);
873 	regs = ap->regs;
874 
875 	board_idx = ap->board_idx;
876 
877 	/*
878 	 * aman@sgi.com - its useful to do a NIC reset here to
879 	 * address the `Firmware not running' problem subsequent
880 	 * to any crashes involving the NIC
881 	 */
882 	writel(HW_RESET | (HW_RESET << 24), &regs->HostCtrl);
883 	readl(&regs->HostCtrl);		/* PCI write posting */
884 	udelay(5);
885 
886 	/*
887 	 * Don't access any other registers before this point!
888 	 */
889 #ifdef __BIG_ENDIAN
890 	/*
891 	 * This will most likely need BYTE_SWAP once we switch
892 	 * to using __raw_writel()
893 	 */
894 	writel((WORD_SWAP | CLR_INT | ((WORD_SWAP | CLR_INT) << 24)),
895 	       &regs->HostCtrl);
896 #else
897 	writel((CLR_INT | WORD_SWAP | ((CLR_INT | WORD_SWAP) << 24)),
898 	       &regs->HostCtrl);
899 #endif
900 	readl(&regs->HostCtrl);		/* PCI write posting */
901 
902 	/*
903 	 * Stop the NIC CPU and clear pending interrupts
904 	 */
905 	writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
906 	readl(&regs->CpuCtrl);		/* PCI write posting */
907 	writel(0, &regs->Mb0Lo);
908 
909 	tig_ver = readl(&regs->HostCtrl) >> 28;
910 
911 	switch(tig_ver){
912 #ifndef CONFIG_ACENIC_OMIT_TIGON_I
913 	case 4:
914 	case 5:
915 		printk(KERN_INFO "  Tigon I  (Rev. %i), Firmware: %i.%i.%i, ",
916 		       tig_ver, ap->firmware_major, ap->firmware_minor,
917 		       ap->firmware_fix);
918 		writel(0, &regs->LocalCtrl);
919 		ap->version = 1;
920 		ap->tx_ring_entries = TIGON_I_TX_RING_ENTRIES;
921 		break;
922 #endif
923 	case 6:
924 		printk(KERN_INFO "  Tigon II (Rev. %i), Firmware: %i.%i.%i, ",
925 		       tig_ver, ap->firmware_major, ap->firmware_minor,
926 		       ap->firmware_fix);
927 		writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
928 		readl(&regs->CpuBCtrl);		/* PCI write posting */
929 		/*
930 		 * The SRAM bank size does _not_ indicate the amount
931 		 * of memory on the card, it controls the _bank_ size!
932 		 * Ie. a 1MB AceNIC will have two banks of 512KB.
933 		 */
934 		writel(SRAM_BANK_512K, &regs->LocalCtrl);
935 		writel(SYNC_SRAM_TIMING, &regs->MiscCfg);
936 		ap->version = 2;
937 		ap->tx_ring_entries = MAX_TX_RING_ENTRIES;
938 		break;
939 	default:
940 		printk(KERN_WARNING "  Unsupported Tigon version detected "
941 		       "(%i)\n", tig_ver);
942 		ecode = -ENODEV;
943 		goto init_error;
944 	}
945 
946 	/*
947 	 * ModeStat _must_ be set after the SRAM settings as this change
948 	 * seems to corrupt the ModeStat and possible other registers.
949 	 * The SRAM settings survive resets and setting it to the same
950 	 * value a second time works as well. This is what caused the
951 	 * `Firmware not running' problem on the Tigon II.
952 	 */
953 #ifdef __BIG_ENDIAN
954 	writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL | ACE_BYTE_SWAP_BD |
955 	       ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
956 #else
957 	writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL |
958 	       ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
959 #endif
960 	readl(&regs->ModeStat);		/* PCI write posting */
961 
962 	mac1 = 0;
963 	for(i = 0; i < 4; i++) {
964 		int t;
965 
966 		mac1 = mac1 << 8;
967 		t = read_eeprom_byte(dev, 0x8c+i);
968 		if (t < 0) {
969 			ecode = -EIO;
970 			goto init_error;
971 		} else
972 			mac1 |= (t & 0xff);
973 	}
974 	mac2 = 0;
975 	for(i = 4; i < 8; i++) {
976 		int t;
977 
978 		mac2 = mac2 << 8;
979 		t = read_eeprom_byte(dev, 0x8c+i);
980 		if (t < 0) {
981 			ecode = -EIO;
982 			goto init_error;
983 		} else
984 			mac2 |= (t & 0xff);
985 	}
986 
987 	writel(mac1, &regs->MacAddrHi);
988 	writel(mac2, &regs->MacAddrLo);
989 
990 	dev->dev_addr[0] = (mac1 >> 8) & 0xff;
991 	dev->dev_addr[1] = mac1 & 0xff;
992 	dev->dev_addr[2] = (mac2 >> 24) & 0xff;
993 	dev->dev_addr[3] = (mac2 >> 16) & 0xff;
994 	dev->dev_addr[4] = (mac2 >> 8) & 0xff;
995 	dev->dev_addr[5] = mac2 & 0xff;
996 
997 	printk("MAC: %pM\n", dev->dev_addr);
998 
999 	/*
1000 	 * Looks like this is necessary to deal with on all architectures,
1001 	 * even this %$#%$# N440BX Intel based thing doesn't get it right.
1002 	 * Ie. having two NICs in the machine, one will have the cache
1003 	 * line set at boot time, the other will not.
1004 	 */
1005 	pdev = ap->pdev;
1006 	pci_read_config_byte(pdev, PCI_CACHE_LINE_SIZE, &cache_size);
1007 	cache_size <<= 2;
1008 	if (cache_size != SMP_CACHE_BYTES) {
1009 		printk(KERN_INFO "  PCI cache line size set incorrectly "
1010 		       "(%i bytes) by BIOS/FW, ", cache_size);
1011 		if (cache_size > SMP_CACHE_BYTES)
1012 			printk("expecting %i\n", SMP_CACHE_BYTES);
1013 		else {
1014 			printk("correcting to %i\n", SMP_CACHE_BYTES);
1015 			pci_write_config_byte(pdev, PCI_CACHE_LINE_SIZE,
1016 					      SMP_CACHE_BYTES >> 2);
1017 		}
1018 	}
1019 
1020 	pci_state = readl(&regs->PciState);
1021 	printk(KERN_INFO "  PCI bus width: %i bits, speed: %iMHz, "
1022 	       "latency: %i clks\n",
1023 	       	(pci_state & PCI_32BIT) ? 32 : 64,
1024 		(pci_state & PCI_66MHZ) ? 66 : 33,
1025 		ap->pci_latency);
1026 
1027 	/*
1028 	 * Set the max DMA transfer size. Seems that for most systems
1029 	 * the performance is better when no MAX parameter is
1030 	 * set. However for systems enabling PCI write and invalidate,
1031 	 * DMA writes must be set to the L1 cache line size to get
1032 	 * optimal performance.
1033 	 *
1034 	 * The default is now to turn the PCI write and invalidate off
1035 	 * - that is what Alteon does for NT.
1036 	 */
1037 	tmp = READ_CMD_MEM | WRITE_CMD_MEM;
1038 	if (ap->version >= 2) {
1039 		tmp |= (MEM_READ_MULTIPLE | (pci_state & PCI_66MHZ));
1040 		/*
1041 		 * Tuning parameters only supported for 8 cards
1042 		 */
1043 		if (board_idx == BOARD_IDX_OVERFLOW ||
1044 		    dis_pci_mem_inval[board_idx]) {
1045 			if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1046 				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1047 				pci_write_config_word(pdev, PCI_COMMAND,
1048 						      ap->pci_command);
1049 				printk(KERN_INFO "  Disabling PCI memory "
1050 				       "write and invalidate\n");
1051 			}
1052 		} else if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1053 			printk(KERN_INFO "  PCI memory write & invalidate "
1054 			       "enabled by BIOS, enabling counter measures\n");
1055 
1056 			switch(SMP_CACHE_BYTES) {
1057 			case 16:
1058 				tmp |= DMA_WRITE_MAX_16;
1059 				break;
1060 			case 32:
1061 				tmp |= DMA_WRITE_MAX_32;
1062 				break;
1063 			case 64:
1064 				tmp |= DMA_WRITE_MAX_64;
1065 				break;
1066 			case 128:
1067 				tmp |= DMA_WRITE_MAX_128;
1068 				break;
1069 			default:
1070 				printk(KERN_INFO "  Cache line size %i not "
1071 				       "supported, PCI write and invalidate "
1072 				       "disabled\n", SMP_CACHE_BYTES);
1073 				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1074 				pci_write_config_word(pdev, PCI_COMMAND,
1075 						      ap->pci_command);
1076 			}
1077 		}
1078 	}
1079 
1080 #ifdef __sparc__
1081 	/*
1082 	 * On this platform, we know what the best dma settings
1083 	 * are.  We use 64-byte maximum bursts, because if we
1084 	 * burst larger than the cache line size (or even cross
1085 	 * a 64byte boundary in a single burst) the UltraSparc
1086 	 * PCI controller will disconnect at 64-byte multiples.
1087 	 *
1088 	 * Read-multiple will be properly enabled above, and when
1089 	 * set will give the PCI controller proper hints about
1090 	 * prefetching.
1091 	 */
1092 	tmp &= ~DMA_READ_WRITE_MASK;
1093 	tmp |= DMA_READ_MAX_64;
1094 	tmp |= DMA_WRITE_MAX_64;
1095 #endif
1096 #ifdef __alpha__
1097 	tmp &= ~DMA_READ_WRITE_MASK;
1098 	tmp |= DMA_READ_MAX_128;
1099 	/*
1100 	 * All the docs say MUST NOT. Well, I did.
1101 	 * Nothing terrible happens, if we load wrong size.
1102 	 * Bit w&i still works better!
1103 	 */
1104 	tmp |= DMA_WRITE_MAX_128;
1105 #endif
1106 	writel(tmp, &regs->PciState);
1107 
1108 #if 0
1109 	/*
1110 	 * The Host PCI bus controller driver has to set FBB.
1111 	 * If all devices on that PCI bus support FBB, then the controller
1112 	 * can enable FBB support in the Host PCI Bus controller (or on
1113 	 * the PCI-PCI bridge if that applies).
1114 	 * -ggg
1115 	 */
1116 	/*
1117 	 * I have received reports from people having problems when this
1118 	 * bit is enabled.
1119 	 */
1120 	if (!(ap->pci_command & PCI_COMMAND_FAST_BACK)) {
1121 		printk(KERN_INFO "  Enabling PCI Fast Back to Back\n");
1122 		ap->pci_command |= PCI_COMMAND_FAST_BACK;
1123 		pci_write_config_word(pdev, PCI_COMMAND, ap->pci_command);
1124 	}
1125 #endif
1126 
1127 	/*
1128 	 * Configure DMA attributes.
1129 	 */
1130 	if (!dma_set_mask(&pdev->dev, DMA_BIT_MASK(64))) {
1131 		ap->pci_using_dac = 1;
1132 	} else if (!dma_set_mask(&pdev->dev, DMA_BIT_MASK(32))) {
1133 		ap->pci_using_dac = 0;
1134 	} else {
1135 		ecode = -ENODEV;
1136 		goto init_error;
1137 	}
1138 
1139 	/*
1140 	 * Initialize the generic info block and the command+event rings
1141 	 * and the control blocks for the transmit and receive rings
1142 	 * as they need to be setup once and for all.
1143 	 */
1144 	if (!(info = dma_alloc_coherent(&ap->pdev->dev, sizeof(struct ace_info),
1145 					&ap->info_dma, GFP_KERNEL))) {
1146 		ecode = -EAGAIN;
1147 		goto init_error;
1148 	}
1149 	ap->info = info;
1150 
1151 	/*
1152 	 * Get the memory for the skb rings.
1153 	 */
1154 	if (!(ap->skb = kzalloc(sizeof(struct ace_skb), GFP_KERNEL))) {
1155 		ecode = -EAGAIN;
1156 		goto init_error;
1157 	}
1158 
1159 	ecode = request_irq(pdev->irq, ace_interrupt, IRQF_SHARED,
1160 			    DRV_NAME, dev);
1161 	if (ecode) {
1162 		printk(KERN_WARNING "%s: Requested IRQ %d is busy\n",
1163 		       DRV_NAME, pdev->irq);
1164 		goto init_error;
1165 	} else
1166 		dev->irq = pdev->irq;
1167 
1168 #ifdef INDEX_DEBUG
1169 	spin_lock_init(&ap->debug_lock);
1170 	ap->last_tx = ACE_TX_RING_ENTRIES(ap) - 1;
1171 	ap->last_std_rx = 0;
1172 	ap->last_mini_rx = 0;
1173 #endif
1174 
1175 	ecode = ace_load_firmware(dev);
1176 	if (ecode)
1177 		goto init_error;
1178 
1179 	ap->fw_running = 0;
1180 
1181 	tmp_ptr = ap->info_dma;
1182 	writel(tmp_ptr >> 32, &regs->InfoPtrHi);
1183 	writel(tmp_ptr & 0xffffffff, &regs->InfoPtrLo);
1184 
1185 	memset(ap->evt_ring, 0, EVT_RING_ENTRIES * sizeof(struct event));
1186 
1187 	set_aceaddr(&info->evt_ctrl.rngptr, ap->evt_ring_dma);
1188 	info->evt_ctrl.flags = 0;
1189 
1190 	*(ap->evt_prd) = 0;
1191 	wmb();
1192 	set_aceaddr(&info->evt_prd_ptr, ap->evt_prd_dma);
1193 	writel(0, &regs->EvtCsm);
1194 
1195 	set_aceaddr(&info->cmd_ctrl.rngptr, 0x100);
1196 	info->cmd_ctrl.flags = 0;
1197 	info->cmd_ctrl.max_len = 0;
1198 
1199 	for (i = 0; i < CMD_RING_ENTRIES; i++)
1200 		writel(0, &regs->CmdRng[i]);
1201 
1202 	writel(0, &regs->CmdPrd);
1203 	writel(0, &regs->CmdCsm);
1204 
1205 	tmp_ptr = ap->info_dma;
1206 	tmp_ptr += (unsigned long) &(((struct ace_info *)0)->s.stats);
1207 	set_aceaddr(&info->stats2_ptr, (dma_addr_t) tmp_ptr);
1208 
1209 	set_aceaddr(&info->rx_std_ctrl.rngptr, ap->rx_ring_base_dma);
1210 	info->rx_std_ctrl.max_len = ACE_STD_BUFSIZE;
1211 	info->rx_std_ctrl.flags =
1212 	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1213 
1214 	memset(ap->rx_std_ring, 0,
1215 	       RX_STD_RING_ENTRIES * sizeof(struct rx_desc));
1216 
1217 	for (i = 0; i < RX_STD_RING_ENTRIES; i++)
1218 		ap->rx_std_ring[i].flags = BD_FLG_TCP_UDP_SUM;
1219 
1220 	ap->rx_std_skbprd = 0;
1221 	atomic_set(&ap->cur_rx_bufs, 0);
1222 
1223 	set_aceaddr(&info->rx_jumbo_ctrl.rngptr,
1224 		    (ap->rx_ring_base_dma +
1225 		     (sizeof(struct rx_desc) * RX_STD_RING_ENTRIES)));
1226 	info->rx_jumbo_ctrl.max_len = 0;
1227 	info->rx_jumbo_ctrl.flags =
1228 	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1229 
1230 	memset(ap->rx_jumbo_ring, 0,
1231 	       RX_JUMBO_RING_ENTRIES * sizeof(struct rx_desc));
1232 
1233 	for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++)
1234 		ap->rx_jumbo_ring[i].flags = BD_FLG_TCP_UDP_SUM | BD_FLG_JUMBO;
1235 
1236 	ap->rx_jumbo_skbprd = 0;
1237 	atomic_set(&ap->cur_jumbo_bufs, 0);
1238 
1239 	memset(ap->rx_mini_ring, 0,
1240 	       RX_MINI_RING_ENTRIES * sizeof(struct rx_desc));
1241 
1242 	if (ap->version >= 2) {
1243 		set_aceaddr(&info->rx_mini_ctrl.rngptr,
1244 			    (ap->rx_ring_base_dma +
1245 			     (sizeof(struct rx_desc) *
1246 			      (RX_STD_RING_ENTRIES +
1247 			       RX_JUMBO_RING_ENTRIES))));
1248 		info->rx_mini_ctrl.max_len = ACE_MINI_SIZE;
1249 		info->rx_mini_ctrl.flags =
1250 		  RCB_FLG_TCP_UDP_SUM|RCB_FLG_NO_PSEUDO_HDR|RCB_FLG_VLAN_ASSIST;
1251 
1252 		for (i = 0; i < RX_MINI_RING_ENTRIES; i++)
1253 			ap->rx_mini_ring[i].flags =
1254 				BD_FLG_TCP_UDP_SUM | BD_FLG_MINI;
1255 	} else {
1256 		set_aceaddr(&info->rx_mini_ctrl.rngptr, 0);
1257 		info->rx_mini_ctrl.flags = RCB_FLG_RNG_DISABLE;
1258 		info->rx_mini_ctrl.max_len = 0;
1259 	}
1260 
1261 	ap->rx_mini_skbprd = 0;
1262 	atomic_set(&ap->cur_mini_bufs, 0);
1263 
1264 	set_aceaddr(&info->rx_return_ctrl.rngptr,
1265 		    (ap->rx_ring_base_dma +
1266 		     (sizeof(struct rx_desc) *
1267 		      (RX_STD_RING_ENTRIES +
1268 		       RX_JUMBO_RING_ENTRIES +
1269 		       RX_MINI_RING_ENTRIES))));
1270 	info->rx_return_ctrl.flags = 0;
1271 	info->rx_return_ctrl.max_len = RX_RETURN_RING_ENTRIES;
1272 
1273 	memset(ap->rx_return_ring, 0,
1274 	       RX_RETURN_RING_ENTRIES * sizeof(struct rx_desc));
1275 
1276 	set_aceaddr(&info->rx_ret_prd_ptr, ap->rx_ret_prd_dma);
1277 	*(ap->rx_ret_prd) = 0;
1278 
1279 	writel(TX_RING_BASE, &regs->WinBase);
1280 
1281 	if (ACE_IS_TIGON_I(ap)) {
1282 		ap->tx_ring = (__force struct tx_desc *) regs->Window;
1283 		for (i = 0; i < (TIGON_I_TX_RING_ENTRIES
1284 				 * sizeof(struct tx_desc)) / sizeof(u32); i++)
1285 			writel(0, (__force void __iomem *)ap->tx_ring  + i * 4);
1286 
1287 		set_aceaddr(&info->tx_ctrl.rngptr, TX_RING_BASE);
1288 	} else {
1289 		memset(ap->tx_ring, 0,
1290 		       MAX_TX_RING_ENTRIES * sizeof(struct tx_desc));
1291 
1292 		set_aceaddr(&info->tx_ctrl.rngptr, ap->tx_ring_dma);
1293 	}
1294 
1295 	info->tx_ctrl.max_len = ACE_TX_RING_ENTRIES(ap);
1296 	tmp = RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1297 
1298 	/*
1299 	 * The Tigon I does not like having the TX ring in host memory ;-(
1300 	 */
1301 	if (!ACE_IS_TIGON_I(ap))
1302 		tmp |= RCB_FLG_TX_HOST_RING;
1303 #if TX_COAL_INTS_ONLY
1304 	tmp |= RCB_FLG_COAL_INT_ONLY;
1305 #endif
1306 	info->tx_ctrl.flags = tmp;
1307 
1308 	set_aceaddr(&info->tx_csm_ptr, ap->tx_csm_dma);
1309 
1310 	/*
1311 	 * Potential item for tuning parameter
1312 	 */
1313 #if 0 /* NO */
1314 	writel(DMA_THRESH_16W, &regs->DmaReadCfg);
1315 	writel(DMA_THRESH_16W, &regs->DmaWriteCfg);
1316 #else
1317 	writel(DMA_THRESH_8W, &regs->DmaReadCfg);
1318 	writel(DMA_THRESH_8W, &regs->DmaWriteCfg);
1319 #endif
1320 
1321 	writel(0, &regs->MaskInt);
1322 	writel(1, &regs->IfIdx);
1323 #if 0
1324 	/*
1325 	 * McKinley boxes do not like us fiddling with AssistState
1326 	 * this early
1327 	 */
1328 	writel(1, &regs->AssistState);
1329 #endif
1330 
1331 	writel(DEF_STAT, &regs->TuneStatTicks);
1332 	writel(DEF_TRACE, &regs->TuneTrace);
1333 
1334 	ace_set_rxtx_parms(dev, 0);
1335 
1336 	if (board_idx == BOARD_IDX_OVERFLOW) {
1337 		printk(KERN_WARNING "%s: more than %i NICs detected, "
1338 		       "ignoring module parameters!\n",
1339 		       ap->name, ACE_MAX_MOD_PARMS);
1340 	} else if (board_idx >= 0) {
1341 		if (tx_coal_tick[board_idx])
1342 			writel(tx_coal_tick[board_idx],
1343 			       &regs->TuneTxCoalTicks);
1344 		if (max_tx_desc[board_idx])
1345 			writel(max_tx_desc[board_idx], &regs->TuneMaxTxDesc);
1346 
1347 		if (rx_coal_tick[board_idx])
1348 			writel(rx_coal_tick[board_idx],
1349 			       &regs->TuneRxCoalTicks);
1350 		if (max_rx_desc[board_idx])
1351 			writel(max_rx_desc[board_idx], &regs->TuneMaxRxDesc);
1352 
1353 		if (trace[board_idx])
1354 			writel(trace[board_idx], &regs->TuneTrace);
1355 
1356 		if ((tx_ratio[board_idx] > 0) && (tx_ratio[board_idx] < 64))
1357 			writel(tx_ratio[board_idx], &regs->TxBufRat);
1358 	}
1359 
1360 	/*
1361 	 * Default link parameters
1362 	 */
1363 	tmp = LNK_ENABLE | LNK_FULL_DUPLEX | LNK_1000MB | LNK_100MB |
1364 		LNK_10MB | LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL | LNK_NEGOTIATE;
1365 	if(ap->version >= 2)
1366 		tmp |= LNK_TX_FLOW_CTL_Y;
1367 
1368 	/*
1369 	 * Override link default parameters
1370 	 */
1371 	if ((board_idx >= 0) && link_state[board_idx]) {
1372 		int option = link_state[board_idx];
1373 
1374 		tmp = LNK_ENABLE;
1375 
1376 		if (option & 0x01) {
1377 			printk(KERN_INFO "%s: Setting half duplex link\n",
1378 			       ap->name);
1379 			tmp &= ~LNK_FULL_DUPLEX;
1380 		}
1381 		if (option & 0x02)
1382 			tmp &= ~LNK_NEGOTIATE;
1383 		if (option & 0x10)
1384 			tmp |= LNK_10MB;
1385 		if (option & 0x20)
1386 			tmp |= LNK_100MB;
1387 		if (option & 0x40)
1388 			tmp |= LNK_1000MB;
1389 		if ((option & 0x70) == 0) {
1390 			printk(KERN_WARNING "%s: No media speed specified, "
1391 			       "forcing auto negotiation\n", ap->name);
1392 			tmp |= LNK_NEGOTIATE | LNK_1000MB |
1393 				LNK_100MB | LNK_10MB;
1394 		}
1395 		if ((option & 0x100) == 0)
1396 			tmp |= LNK_NEG_FCTL;
1397 		else
1398 			printk(KERN_INFO "%s: Disabling flow control "
1399 			       "negotiation\n", ap->name);
1400 		if (option & 0x200)
1401 			tmp |= LNK_RX_FLOW_CTL_Y;
1402 		if ((option & 0x400) && (ap->version >= 2)) {
1403 			printk(KERN_INFO "%s: Enabling TX flow control\n",
1404 			       ap->name);
1405 			tmp |= LNK_TX_FLOW_CTL_Y;
1406 		}
1407 	}
1408 
1409 	ap->link = tmp;
1410 	writel(tmp, &regs->TuneLink);
1411 	if (ap->version >= 2)
1412 		writel(tmp, &regs->TuneFastLink);
1413 
1414 	writel(ap->firmware_start, &regs->Pc);
1415 
1416 	writel(0, &regs->Mb0Lo);
1417 
1418 	/*
1419 	 * Set tx_csm before we start receiving interrupts, otherwise
1420 	 * the interrupt handler might think it is supposed to process
1421 	 * tx ints before we are up and running, which may cause a null
1422 	 * pointer access in the int handler.
1423 	 */
1424 	ap->cur_rx = 0;
1425 	ap->tx_prd = *(ap->tx_csm) = ap->tx_ret_csm = 0;
1426 
1427 	wmb();
1428 	ace_set_txprd(regs, ap, 0);
1429 	writel(0, &regs->RxRetCsm);
1430 
1431 	/*
1432 	 * Enable DMA engine now.
1433 	 * If we do this sooner, Mckinley box pukes.
1434 	 * I assume it's because Tigon II DMA engine wants to check
1435 	 * *something* even before the CPU is started.
1436 	 */
1437 	writel(1, &regs->AssistState);  /* enable DMA */
1438 
1439 	/*
1440 	 * Start the NIC CPU
1441 	 */
1442 	writel(readl(&regs->CpuCtrl) & ~(CPU_HALT|CPU_TRACE), &regs->CpuCtrl);
1443 	readl(&regs->CpuCtrl);
1444 
1445 	/*
1446 	 * Wait for the firmware to spin up - max 3 seconds.
1447 	 */
1448 	myjif = jiffies + 3 * HZ;
1449 	while (time_before(jiffies, myjif) && !ap->fw_running)
1450 		cpu_relax();
1451 
1452 	if (!ap->fw_running) {
1453 		printk(KERN_ERR "%s: Firmware NOT running!\n", ap->name);
1454 
1455 		ace_dump_trace(ap);
1456 		writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
1457 		readl(&regs->CpuCtrl);
1458 
1459 		/* aman@sgi.com - account for badly behaving firmware/NIC:
1460 		 * - have observed that the NIC may continue to generate
1461 		 *   interrupts for some reason; attempt to stop it - halt
1462 		 *   second CPU for Tigon II cards, and also clear Mb0
1463 		 * - if we're a module, we'll fail to load if this was
1464 		 *   the only GbE card in the system => if the kernel does
1465 		 *   see an interrupt from the NIC, code to handle it is
1466 		 *   gone and OOps! - so free_irq also
1467 		 */
1468 		if (ap->version >= 2)
1469 			writel(readl(&regs->CpuBCtrl) | CPU_HALT,
1470 			       &regs->CpuBCtrl);
1471 		writel(0, &regs->Mb0Lo);
1472 		readl(&regs->Mb0Lo);
1473 
1474 		ecode = -EBUSY;
1475 		goto init_error;
1476 	}
1477 
1478 	/*
1479 	 * We load the ring here as there seem to be no way to tell the
1480 	 * firmware to wipe the ring without re-initializing it.
1481 	 */
1482 	if (!test_and_set_bit(0, &ap->std_refill_busy))
1483 		ace_load_std_rx_ring(dev, RX_RING_SIZE);
1484 	else
1485 		printk(KERN_ERR "%s: Someone is busy refilling the RX ring\n",
1486 		       ap->name);
1487 	if (ap->version >= 2) {
1488 		if (!test_and_set_bit(0, &ap->mini_refill_busy))
1489 			ace_load_mini_rx_ring(dev, RX_MINI_SIZE);
1490 		else
1491 			printk(KERN_ERR "%s: Someone is busy refilling "
1492 			       "the RX mini ring\n", ap->name);
1493 	}
1494 	return 0;
1495 
1496  init_error:
1497 	ace_init_cleanup(dev);
1498 	return ecode;
1499 }
1500 
1501 
1502 static void ace_set_rxtx_parms(struct net_device *dev, int jumbo)
1503 {
1504 	struct ace_private *ap = netdev_priv(dev);
1505 	struct ace_regs __iomem *regs = ap->regs;
1506 	int board_idx = ap->board_idx;
1507 
1508 	if (board_idx >= 0) {
1509 		if (!jumbo) {
1510 			if (!tx_coal_tick[board_idx])
1511 				writel(DEF_TX_COAL, &regs->TuneTxCoalTicks);
1512 			if (!max_tx_desc[board_idx])
1513 				writel(DEF_TX_MAX_DESC, &regs->TuneMaxTxDesc);
1514 			if (!rx_coal_tick[board_idx])
1515 				writel(DEF_RX_COAL, &regs->TuneRxCoalTicks);
1516 			if (!max_rx_desc[board_idx])
1517 				writel(DEF_RX_MAX_DESC, &regs->TuneMaxRxDesc);
1518 			if (!tx_ratio[board_idx])
1519 				writel(DEF_TX_RATIO, &regs->TxBufRat);
1520 		} else {
1521 			if (!tx_coal_tick[board_idx])
1522 				writel(DEF_JUMBO_TX_COAL,
1523 				       &regs->TuneTxCoalTicks);
1524 			if (!max_tx_desc[board_idx])
1525 				writel(DEF_JUMBO_TX_MAX_DESC,
1526 				       &regs->TuneMaxTxDesc);
1527 			if (!rx_coal_tick[board_idx])
1528 				writel(DEF_JUMBO_RX_COAL,
1529 				       &regs->TuneRxCoalTicks);
1530 			if (!max_rx_desc[board_idx])
1531 				writel(DEF_JUMBO_RX_MAX_DESC,
1532 				       &regs->TuneMaxRxDesc);
1533 			if (!tx_ratio[board_idx])
1534 				writel(DEF_JUMBO_TX_RATIO, &regs->TxBufRat);
1535 		}
1536 	}
1537 }
1538 
1539 
1540 static void ace_watchdog(struct net_device *data, unsigned int txqueue)
1541 {
1542 	struct net_device *dev = data;
1543 	struct ace_private *ap = netdev_priv(dev);
1544 	struct ace_regs __iomem *regs = ap->regs;
1545 
1546 	/*
1547 	 * We haven't received a stats update event for more than 2.5
1548 	 * seconds and there is data in the transmit queue, thus we
1549 	 * assume the card is stuck.
1550 	 */
1551 	if (*ap->tx_csm != ap->tx_ret_csm) {
1552 		printk(KERN_WARNING "%s: Transmitter is stuck, %08x\n",
1553 		       dev->name, (unsigned int)readl(&regs->HostCtrl));
1554 		/* This can happen due to ieee flow control. */
1555 	} else {
1556 		printk(KERN_DEBUG "%s: BUG... transmitter died. Kicking it.\n",
1557 		       dev->name);
1558 #if 0
1559 		netif_wake_queue(dev);
1560 #endif
1561 	}
1562 }
1563 
1564 
1565 static void ace_tasklet(unsigned long arg)
1566 {
1567 	struct net_device *dev = (struct net_device *) arg;
1568 	struct ace_private *ap = netdev_priv(dev);
1569 	int cur_size;
1570 
1571 	cur_size = atomic_read(&ap->cur_rx_bufs);
1572 	if ((cur_size < RX_LOW_STD_THRES) &&
1573 	    !test_and_set_bit(0, &ap->std_refill_busy)) {
1574 #ifdef DEBUG
1575 		printk("refilling buffers (current %i)\n", cur_size);
1576 #endif
1577 		ace_load_std_rx_ring(dev, RX_RING_SIZE - cur_size);
1578 	}
1579 
1580 	if (ap->version >= 2) {
1581 		cur_size = atomic_read(&ap->cur_mini_bufs);
1582 		if ((cur_size < RX_LOW_MINI_THRES) &&
1583 		    !test_and_set_bit(0, &ap->mini_refill_busy)) {
1584 #ifdef DEBUG
1585 			printk("refilling mini buffers (current %i)\n",
1586 			       cur_size);
1587 #endif
1588 			ace_load_mini_rx_ring(dev, RX_MINI_SIZE - cur_size);
1589 		}
1590 	}
1591 
1592 	cur_size = atomic_read(&ap->cur_jumbo_bufs);
1593 	if (ap->jumbo && (cur_size < RX_LOW_JUMBO_THRES) &&
1594 	    !test_and_set_bit(0, &ap->jumbo_refill_busy)) {
1595 #ifdef DEBUG
1596 		printk("refilling jumbo buffers (current %i)\n", cur_size);
1597 #endif
1598 		ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE - cur_size);
1599 	}
1600 	ap->tasklet_pending = 0;
1601 }
1602 
1603 
1604 /*
1605  * Copy the contents of the NIC's trace buffer to kernel memory.
1606  */
1607 static void ace_dump_trace(struct ace_private *ap)
1608 {
1609 #if 0
1610 	if (!ap->trace_buf)
1611 		if (!(ap->trace_buf = kmalloc(ACE_TRACE_SIZE, GFP_KERNEL)))
1612 		    return;
1613 #endif
1614 }
1615 
1616 
1617 /*
1618  * Load the standard rx ring.
1619  *
1620  * Loading rings is safe without holding the spin lock since this is
1621  * done only before the device is enabled, thus no interrupts are
1622  * generated and by the interrupt handler/tasklet handler.
1623  */
1624 static void ace_load_std_rx_ring(struct net_device *dev, int nr_bufs)
1625 {
1626 	struct ace_private *ap = netdev_priv(dev);
1627 	struct ace_regs __iomem *regs = ap->regs;
1628 	short i, idx;
1629 
1630 
1631 	prefetchw(&ap->cur_rx_bufs);
1632 
1633 	idx = ap->rx_std_skbprd;
1634 
1635 	for (i = 0; i < nr_bufs; i++) {
1636 		struct sk_buff *skb;
1637 		struct rx_desc *rd;
1638 		dma_addr_t mapping;
1639 
1640 		skb = netdev_alloc_skb_ip_align(dev, ACE_STD_BUFSIZE);
1641 		if (!skb)
1642 			break;
1643 
1644 		mapping = dma_map_page(&ap->pdev->dev,
1645 				       virt_to_page(skb->data),
1646 				       offset_in_page(skb->data),
1647 				       ACE_STD_BUFSIZE, DMA_FROM_DEVICE);
1648 		ap->skb->rx_std_skbuff[idx].skb = skb;
1649 		dma_unmap_addr_set(&ap->skb->rx_std_skbuff[idx],
1650 				   mapping, mapping);
1651 
1652 		rd = &ap->rx_std_ring[idx];
1653 		set_aceaddr(&rd->addr, mapping);
1654 		rd->size = ACE_STD_BUFSIZE;
1655 		rd->idx = idx;
1656 		idx = (idx + 1) % RX_STD_RING_ENTRIES;
1657 	}
1658 
1659 	if (!i)
1660 		goto error_out;
1661 
1662 	atomic_add(i, &ap->cur_rx_bufs);
1663 	ap->rx_std_skbprd = idx;
1664 
1665 	if (ACE_IS_TIGON_I(ap)) {
1666 		struct cmd cmd;
1667 		cmd.evt = C_SET_RX_PRD_IDX;
1668 		cmd.code = 0;
1669 		cmd.idx = ap->rx_std_skbprd;
1670 		ace_issue_cmd(regs, &cmd);
1671 	} else {
1672 		writel(idx, &regs->RxStdPrd);
1673 		wmb();
1674 	}
1675 
1676  out:
1677 	clear_bit(0, &ap->std_refill_busy);
1678 	return;
1679 
1680  error_out:
1681 	printk(KERN_INFO "Out of memory when allocating "
1682 	       "standard receive buffers\n");
1683 	goto out;
1684 }
1685 
1686 
1687 static void ace_load_mini_rx_ring(struct net_device *dev, int nr_bufs)
1688 {
1689 	struct ace_private *ap = netdev_priv(dev);
1690 	struct ace_regs __iomem *regs = ap->regs;
1691 	short i, idx;
1692 
1693 	prefetchw(&ap->cur_mini_bufs);
1694 
1695 	idx = ap->rx_mini_skbprd;
1696 	for (i = 0; i < nr_bufs; i++) {
1697 		struct sk_buff *skb;
1698 		struct rx_desc *rd;
1699 		dma_addr_t mapping;
1700 
1701 		skb = netdev_alloc_skb_ip_align(dev, ACE_MINI_BUFSIZE);
1702 		if (!skb)
1703 			break;
1704 
1705 		mapping = dma_map_page(&ap->pdev->dev,
1706 				       virt_to_page(skb->data),
1707 				       offset_in_page(skb->data),
1708 				       ACE_MINI_BUFSIZE, DMA_FROM_DEVICE);
1709 		ap->skb->rx_mini_skbuff[idx].skb = skb;
1710 		dma_unmap_addr_set(&ap->skb->rx_mini_skbuff[idx],
1711 				   mapping, mapping);
1712 
1713 		rd = &ap->rx_mini_ring[idx];
1714 		set_aceaddr(&rd->addr, mapping);
1715 		rd->size = ACE_MINI_BUFSIZE;
1716 		rd->idx = idx;
1717 		idx = (idx + 1) % RX_MINI_RING_ENTRIES;
1718 	}
1719 
1720 	if (!i)
1721 		goto error_out;
1722 
1723 	atomic_add(i, &ap->cur_mini_bufs);
1724 
1725 	ap->rx_mini_skbprd = idx;
1726 
1727 	writel(idx, &regs->RxMiniPrd);
1728 	wmb();
1729 
1730  out:
1731 	clear_bit(0, &ap->mini_refill_busy);
1732 	return;
1733  error_out:
1734 	printk(KERN_INFO "Out of memory when allocating "
1735 	       "mini receive buffers\n");
1736 	goto out;
1737 }
1738 
1739 
1740 /*
1741  * Load the jumbo rx ring, this may happen at any time if the MTU
1742  * is changed to a value > 1500.
1743  */
1744 static void ace_load_jumbo_rx_ring(struct net_device *dev, int nr_bufs)
1745 {
1746 	struct ace_private *ap = netdev_priv(dev);
1747 	struct ace_regs __iomem *regs = ap->regs;
1748 	short i, idx;
1749 
1750 	idx = ap->rx_jumbo_skbprd;
1751 
1752 	for (i = 0; i < nr_bufs; i++) {
1753 		struct sk_buff *skb;
1754 		struct rx_desc *rd;
1755 		dma_addr_t mapping;
1756 
1757 		skb = netdev_alloc_skb_ip_align(dev, ACE_JUMBO_BUFSIZE);
1758 		if (!skb)
1759 			break;
1760 
1761 		mapping = dma_map_page(&ap->pdev->dev,
1762 				       virt_to_page(skb->data),
1763 				       offset_in_page(skb->data),
1764 				       ACE_JUMBO_BUFSIZE, DMA_FROM_DEVICE);
1765 		ap->skb->rx_jumbo_skbuff[idx].skb = skb;
1766 		dma_unmap_addr_set(&ap->skb->rx_jumbo_skbuff[idx],
1767 				   mapping, mapping);
1768 
1769 		rd = &ap->rx_jumbo_ring[idx];
1770 		set_aceaddr(&rd->addr, mapping);
1771 		rd->size = ACE_JUMBO_BUFSIZE;
1772 		rd->idx = idx;
1773 		idx = (idx + 1) % RX_JUMBO_RING_ENTRIES;
1774 	}
1775 
1776 	if (!i)
1777 		goto error_out;
1778 
1779 	atomic_add(i, &ap->cur_jumbo_bufs);
1780 	ap->rx_jumbo_skbprd = idx;
1781 
1782 	if (ACE_IS_TIGON_I(ap)) {
1783 		struct cmd cmd;
1784 		cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1785 		cmd.code = 0;
1786 		cmd.idx = ap->rx_jumbo_skbprd;
1787 		ace_issue_cmd(regs, &cmd);
1788 	} else {
1789 		writel(idx, &regs->RxJumboPrd);
1790 		wmb();
1791 	}
1792 
1793  out:
1794 	clear_bit(0, &ap->jumbo_refill_busy);
1795 	return;
1796  error_out:
1797 	if (net_ratelimit())
1798 		printk(KERN_INFO "Out of memory when allocating "
1799 		       "jumbo receive buffers\n");
1800 	goto out;
1801 }
1802 
1803 
1804 /*
1805  * All events are considered to be slow (RX/TX ints do not generate
1806  * events) and are handled here, outside the main interrupt handler,
1807  * to reduce the size of the handler.
1808  */
1809 static u32 ace_handle_event(struct net_device *dev, u32 evtcsm, u32 evtprd)
1810 {
1811 	struct ace_private *ap;
1812 
1813 	ap = netdev_priv(dev);
1814 
1815 	while (evtcsm != evtprd) {
1816 		switch (ap->evt_ring[evtcsm].evt) {
1817 		case E_FW_RUNNING:
1818 			printk(KERN_INFO "%s: Firmware up and running\n",
1819 			       ap->name);
1820 			ap->fw_running = 1;
1821 			wmb();
1822 			break;
1823 		case E_STATS_UPDATED:
1824 			break;
1825 		case E_LNK_STATE:
1826 		{
1827 			u16 code = ap->evt_ring[evtcsm].code;
1828 			switch (code) {
1829 			case E_C_LINK_UP:
1830 			{
1831 				u32 state = readl(&ap->regs->GigLnkState);
1832 				printk(KERN_WARNING "%s: Optical link UP "
1833 				       "(%s Duplex, Flow Control: %s%s)\n",
1834 				       ap->name,
1835 				       state & LNK_FULL_DUPLEX ? "Full":"Half",
1836 				       state & LNK_TX_FLOW_CTL_Y ? "TX " : "",
1837 				       state & LNK_RX_FLOW_CTL_Y ? "RX" : "");
1838 				break;
1839 			}
1840 			case E_C_LINK_DOWN:
1841 				printk(KERN_WARNING "%s: Optical link DOWN\n",
1842 				       ap->name);
1843 				break;
1844 			case E_C_LINK_10_100:
1845 				printk(KERN_WARNING "%s: 10/100BaseT link "
1846 				       "UP\n", ap->name);
1847 				break;
1848 			default:
1849 				printk(KERN_ERR "%s: Unknown optical link "
1850 				       "state %02x\n", ap->name, code);
1851 			}
1852 			break;
1853 		}
1854 		case E_ERROR:
1855 			switch(ap->evt_ring[evtcsm].code) {
1856 			case E_C_ERR_INVAL_CMD:
1857 				printk(KERN_ERR "%s: invalid command error\n",
1858 				       ap->name);
1859 				break;
1860 			case E_C_ERR_UNIMP_CMD:
1861 				printk(KERN_ERR "%s: unimplemented command "
1862 				       "error\n", ap->name);
1863 				break;
1864 			case E_C_ERR_BAD_CFG:
1865 				printk(KERN_ERR "%s: bad config error\n",
1866 				       ap->name);
1867 				break;
1868 			default:
1869 				printk(KERN_ERR "%s: unknown error %02x\n",
1870 				       ap->name, ap->evt_ring[evtcsm].code);
1871 			}
1872 			break;
1873 		case E_RESET_JUMBO_RNG:
1874 		{
1875 			int i;
1876 			for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
1877 				if (ap->skb->rx_jumbo_skbuff[i].skb) {
1878 					ap->rx_jumbo_ring[i].size = 0;
1879 					set_aceaddr(&ap->rx_jumbo_ring[i].addr, 0);
1880 					dev_kfree_skb(ap->skb->rx_jumbo_skbuff[i].skb);
1881 					ap->skb->rx_jumbo_skbuff[i].skb = NULL;
1882 				}
1883 			}
1884 
1885  			if (ACE_IS_TIGON_I(ap)) {
1886  				struct cmd cmd;
1887  				cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1888  				cmd.code = 0;
1889  				cmd.idx = 0;
1890  				ace_issue_cmd(ap->regs, &cmd);
1891  			} else {
1892  				writel(0, &((ap->regs)->RxJumboPrd));
1893  				wmb();
1894  			}
1895 
1896 			ap->jumbo = 0;
1897 			ap->rx_jumbo_skbprd = 0;
1898 			printk(KERN_INFO "%s: Jumbo ring flushed\n",
1899 			       ap->name);
1900 			clear_bit(0, &ap->jumbo_refill_busy);
1901 			break;
1902 		}
1903 		default:
1904 			printk(KERN_ERR "%s: Unhandled event 0x%02x\n",
1905 			       ap->name, ap->evt_ring[evtcsm].evt);
1906 		}
1907 		evtcsm = (evtcsm + 1) % EVT_RING_ENTRIES;
1908 	}
1909 
1910 	return evtcsm;
1911 }
1912 
1913 
1914 static void ace_rx_int(struct net_device *dev, u32 rxretprd, u32 rxretcsm)
1915 {
1916 	struct ace_private *ap = netdev_priv(dev);
1917 	u32 idx;
1918 	int mini_count = 0, std_count = 0;
1919 
1920 	idx = rxretcsm;
1921 
1922 	prefetchw(&ap->cur_rx_bufs);
1923 	prefetchw(&ap->cur_mini_bufs);
1924 
1925 	while (idx != rxretprd) {
1926 		struct ring_info *rip;
1927 		struct sk_buff *skb;
1928 		struct rx_desc *retdesc;
1929 		u32 skbidx;
1930 		int bd_flags, desc_type, mapsize;
1931 		u16 csum;
1932 
1933 
1934 		/* make sure the rx descriptor isn't read before rxretprd */
1935 		if (idx == rxretcsm)
1936 			rmb();
1937 
1938 		retdesc = &ap->rx_return_ring[idx];
1939 		skbidx = retdesc->idx;
1940 		bd_flags = retdesc->flags;
1941 		desc_type = bd_flags & (BD_FLG_JUMBO | BD_FLG_MINI);
1942 
1943 		switch(desc_type) {
1944 			/*
1945 			 * Normal frames do not have any flags set
1946 			 *
1947 			 * Mini and normal frames arrive frequently,
1948 			 * so use a local counter to avoid doing
1949 			 * atomic operations for each packet arriving.
1950 			 */
1951 		case 0:
1952 			rip = &ap->skb->rx_std_skbuff[skbidx];
1953 			mapsize = ACE_STD_BUFSIZE;
1954 			std_count++;
1955 			break;
1956 		case BD_FLG_JUMBO:
1957 			rip = &ap->skb->rx_jumbo_skbuff[skbidx];
1958 			mapsize = ACE_JUMBO_BUFSIZE;
1959 			atomic_dec(&ap->cur_jumbo_bufs);
1960 			break;
1961 		case BD_FLG_MINI:
1962 			rip = &ap->skb->rx_mini_skbuff[skbidx];
1963 			mapsize = ACE_MINI_BUFSIZE;
1964 			mini_count++;
1965 			break;
1966 		default:
1967 			printk(KERN_INFO "%s: unknown frame type (0x%02x) "
1968 			       "returned by NIC\n", dev->name,
1969 			       retdesc->flags);
1970 			goto error;
1971 		}
1972 
1973 		skb = rip->skb;
1974 		rip->skb = NULL;
1975 		dma_unmap_page(&ap->pdev->dev, dma_unmap_addr(rip, mapping),
1976 			       mapsize, DMA_FROM_DEVICE);
1977 		skb_put(skb, retdesc->size);
1978 
1979 		/*
1980 		 * Fly baby, fly!
1981 		 */
1982 		csum = retdesc->tcp_udp_csum;
1983 
1984 		skb->protocol = eth_type_trans(skb, dev);
1985 
1986 		/*
1987 		 * Instead of forcing the poor tigon mips cpu to calculate
1988 		 * pseudo hdr checksum, we do this ourselves.
1989 		 */
1990 		if (bd_flags & BD_FLG_TCP_UDP_SUM) {
1991 			skb->csum = htons(csum);
1992 			skb->ip_summed = CHECKSUM_COMPLETE;
1993 		} else {
1994 			skb_checksum_none_assert(skb);
1995 		}
1996 
1997 		/* send it up */
1998 		if ((bd_flags & BD_FLG_VLAN_TAG))
1999 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), retdesc->vlan);
2000 		netif_rx(skb);
2001 
2002 		dev->stats.rx_packets++;
2003 		dev->stats.rx_bytes += retdesc->size;
2004 
2005 		idx = (idx + 1) % RX_RETURN_RING_ENTRIES;
2006 	}
2007 
2008 	atomic_sub(std_count, &ap->cur_rx_bufs);
2009 	if (!ACE_IS_TIGON_I(ap))
2010 		atomic_sub(mini_count, &ap->cur_mini_bufs);
2011 
2012  out:
2013 	/*
2014 	 * According to the documentation RxRetCsm is obsolete with
2015 	 * the 12.3.x Firmware - my Tigon I NICs seem to disagree!
2016 	 */
2017 	if (ACE_IS_TIGON_I(ap)) {
2018 		writel(idx, &ap->regs->RxRetCsm);
2019 	}
2020 	ap->cur_rx = idx;
2021 
2022 	return;
2023  error:
2024 	idx = rxretprd;
2025 	goto out;
2026 }
2027 
2028 
2029 static inline void ace_tx_int(struct net_device *dev,
2030 			      u32 txcsm, u32 idx)
2031 {
2032 	struct ace_private *ap = netdev_priv(dev);
2033 
2034 	do {
2035 		struct sk_buff *skb;
2036 		struct tx_ring_info *info;
2037 
2038 		info = ap->skb->tx_skbuff + idx;
2039 		skb = info->skb;
2040 
2041 		if (dma_unmap_len(info, maplen)) {
2042 			dma_unmap_page(&ap->pdev->dev,
2043 				       dma_unmap_addr(info, mapping),
2044 				       dma_unmap_len(info, maplen),
2045 				       DMA_TO_DEVICE);
2046 			dma_unmap_len_set(info, maplen, 0);
2047 		}
2048 
2049 		if (skb) {
2050 			dev->stats.tx_packets++;
2051 			dev->stats.tx_bytes += skb->len;
2052 			dev_consume_skb_irq(skb);
2053 			info->skb = NULL;
2054 		}
2055 
2056 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2057 	} while (idx != txcsm);
2058 
2059 	if (netif_queue_stopped(dev))
2060 		netif_wake_queue(dev);
2061 
2062 	wmb();
2063 	ap->tx_ret_csm = txcsm;
2064 
2065 	/* So... tx_ret_csm is advanced _after_ check for device wakeup.
2066 	 *
2067 	 * We could try to make it before. In this case we would get
2068 	 * the following race condition: hard_start_xmit on other cpu
2069 	 * enters after we advanced tx_ret_csm and fills space,
2070 	 * which we have just freed, so that we make illegal device wakeup.
2071 	 * There is no good way to workaround this (at entry
2072 	 * to ace_start_xmit detects this condition and prevents
2073 	 * ring corruption, but it is not a good workaround.)
2074 	 *
2075 	 * When tx_ret_csm is advanced after, we wake up device _only_
2076 	 * if we really have some space in ring (though the core doing
2077 	 * hard_start_xmit can see full ring for some period and has to
2078 	 * synchronize.) Superb.
2079 	 * BUT! We get another subtle race condition. hard_start_xmit
2080 	 * may think that ring is full between wakeup and advancing
2081 	 * tx_ret_csm and will stop device instantly! It is not so bad.
2082 	 * We are guaranteed that there is something in ring, so that
2083 	 * the next irq will resume transmission. To speedup this we could
2084 	 * mark descriptor, which closes ring with BD_FLG_COAL_NOW
2085 	 * (see ace_start_xmit).
2086 	 *
2087 	 * Well, this dilemma exists in all lock-free devices.
2088 	 * We, following scheme used in drivers by Donald Becker,
2089 	 * select the least dangerous.
2090 	 *							--ANK
2091 	 */
2092 }
2093 
2094 
2095 static irqreturn_t ace_interrupt(int irq, void *dev_id)
2096 {
2097 	struct net_device *dev = (struct net_device *)dev_id;
2098 	struct ace_private *ap = netdev_priv(dev);
2099 	struct ace_regs __iomem *regs = ap->regs;
2100 	u32 idx;
2101 	u32 txcsm, rxretcsm, rxretprd;
2102 	u32 evtcsm, evtprd;
2103 
2104 	/*
2105 	 * In case of PCI shared interrupts or spurious interrupts,
2106 	 * we want to make sure it is actually our interrupt before
2107 	 * spending any time in here.
2108 	 */
2109 	if (!(readl(&regs->HostCtrl) & IN_INT))
2110 		return IRQ_NONE;
2111 
2112 	/*
2113 	 * ACK intr now. Otherwise we will lose updates to rx_ret_prd,
2114 	 * which happened _after_ rxretprd = *ap->rx_ret_prd; but before
2115 	 * writel(0, &regs->Mb0Lo).
2116 	 *
2117 	 * "IRQ avoidance" recommended in docs applies to IRQs served
2118 	 * threads and it is wrong even for that case.
2119 	 */
2120 	writel(0, &regs->Mb0Lo);
2121 	readl(&regs->Mb0Lo);
2122 
2123 	/*
2124 	 * There is no conflict between transmit handling in
2125 	 * start_xmit and receive processing, thus there is no reason
2126 	 * to take a spin lock for RX handling. Wait until we start
2127 	 * working on the other stuff - hey we don't need a spin lock
2128 	 * anymore.
2129 	 */
2130 	rxretprd = *ap->rx_ret_prd;
2131 	rxretcsm = ap->cur_rx;
2132 
2133 	if (rxretprd != rxretcsm)
2134 		ace_rx_int(dev, rxretprd, rxretcsm);
2135 
2136 	txcsm = *ap->tx_csm;
2137 	idx = ap->tx_ret_csm;
2138 
2139 	if (txcsm != idx) {
2140 		/*
2141 		 * If each skb takes only one descriptor this check degenerates
2142 		 * to identity, because new space has just been opened.
2143 		 * But if skbs are fragmented we must check that this index
2144 		 * update releases enough of space, otherwise we just
2145 		 * wait for device to make more work.
2146 		 */
2147 		if (!tx_ring_full(ap, txcsm, ap->tx_prd))
2148 			ace_tx_int(dev, txcsm, idx);
2149 	}
2150 
2151 	evtcsm = readl(&regs->EvtCsm);
2152 	evtprd = *ap->evt_prd;
2153 
2154 	if (evtcsm != evtprd) {
2155 		evtcsm = ace_handle_event(dev, evtcsm, evtprd);
2156 		writel(evtcsm, &regs->EvtCsm);
2157 	}
2158 
2159 	/*
2160 	 * This has to go last in the interrupt handler and run with
2161 	 * the spin lock released ... what lock?
2162 	 */
2163 	if (netif_running(dev)) {
2164 		int cur_size;
2165 		int run_tasklet = 0;
2166 
2167 		cur_size = atomic_read(&ap->cur_rx_bufs);
2168 		if (cur_size < RX_LOW_STD_THRES) {
2169 			if ((cur_size < RX_PANIC_STD_THRES) &&
2170 			    !test_and_set_bit(0, &ap->std_refill_busy)) {
2171 #ifdef DEBUG
2172 				printk("low on std buffers %i\n", cur_size);
2173 #endif
2174 				ace_load_std_rx_ring(dev,
2175 						     RX_RING_SIZE - cur_size);
2176 			} else
2177 				run_tasklet = 1;
2178 		}
2179 
2180 		if (!ACE_IS_TIGON_I(ap)) {
2181 			cur_size = atomic_read(&ap->cur_mini_bufs);
2182 			if (cur_size < RX_LOW_MINI_THRES) {
2183 				if ((cur_size < RX_PANIC_MINI_THRES) &&
2184 				    !test_and_set_bit(0,
2185 						      &ap->mini_refill_busy)) {
2186 #ifdef DEBUG
2187 					printk("low on mini buffers %i\n",
2188 					       cur_size);
2189 #endif
2190 					ace_load_mini_rx_ring(dev,
2191 							      RX_MINI_SIZE - cur_size);
2192 				} else
2193 					run_tasklet = 1;
2194 			}
2195 		}
2196 
2197 		if (ap->jumbo) {
2198 			cur_size = atomic_read(&ap->cur_jumbo_bufs);
2199 			if (cur_size < RX_LOW_JUMBO_THRES) {
2200 				if ((cur_size < RX_PANIC_JUMBO_THRES) &&
2201 				    !test_and_set_bit(0,
2202 						      &ap->jumbo_refill_busy)){
2203 #ifdef DEBUG
2204 					printk("low on jumbo buffers %i\n",
2205 					       cur_size);
2206 #endif
2207 					ace_load_jumbo_rx_ring(dev,
2208 							       RX_JUMBO_SIZE - cur_size);
2209 				} else
2210 					run_tasklet = 1;
2211 			}
2212 		}
2213 		if (run_tasklet && !ap->tasklet_pending) {
2214 			ap->tasklet_pending = 1;
2215 			tasklet_schedule(&ap->ace_tasklet);
2216 		}
2217 	}
2218 
2219 	return IRQ_HANDLED;
2220 }
2221 
2222 static int ace_open(struct net_device *dev)
2223 {
2224 	struct ace_private *ap = netdev_priv(dev);
2225 	struct ace_regs __iomem *regs = ap->regs;
2226 	struct cmd cmd;
2227 
2228 	if (!(ap->fw_running)) {
2229 		printk(KERN_WARNING "%s: Firmware not running!\n", dev->name);
2230 		return -EBUSY;
2231 	}
2232 
2233 	writel(dev->mtu + ETH_HLEN + 4, &regs->IfMtu);
2234 
2235 	cmd.evt = C_CLEAR_STATS;
2236 	cmd.code = 0;
2237 	cmd.idx = 0;
2238 	ace_issue_cmd(regs, &cmd);
2239 
2240 	cmd.evt = C_HOST_STATE;
2241 	cmd.code = C_C_STACK_UP;
2242 	cmd.idx = 0;
2243 	ace_issue_cmd(regs, &cmd);
2244 
2245 	if (ap->jumbo &&
2246 	    !test_and_set_bit(0, &ap->jumbo_refill_busy))
2247 		ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE);
2248 
2249 	if (dev->flags & IFF_PROMISC) {
2250 		cmd.evt = C_SET_PROMISC_MODE;
2251 		cmd.code = C_C_PROMISC_ENABLE;
2252 		cmd.idx = 0;
2253 		ace_issue_cmd(regs, &cmd);
2254 
2255 		ap->promisc = 1;
2256 	}else
2257 		ap->promisc = 0;
2258 	ap->mcast_all = 0;
2259 
2260 #if 0
2261 	cmd.evt = C_LNK_NEGOTIATION;
2262 	cmd.code = 0;
2263 	cmd.idx = 0;
2264 	ace_issue_cmd(regs, &cmd);
2265 #endif
2266 
2267 	netif_start_queue(dev);
2268 
2269 	/*
2270 	 * Setup the bottom half rx ring refill handler
2271 	 */
2272 	tasklet_init(&ap->ace_tasklet, ace_tasklet, (unsigned long)dev);
2273 	return 0;
2274 }
2275 
2276 
2277 static int ace_close(struct net_device *dev)
2278 {
2279 	struct ace_private *ap = netdev_priv(dev);
2280 	struct ace_regs __iomem *regs = ap->regs;
2281 	struct cmd cmd;
2282 	unsigned long flags;
2283 	short i;
2284 
2285 	/*
2286 	 * Without (or before) releasing irq and stopping hardware, this
2287 	 * is an absolute non-sense, by the way. It will be reset instantly
2288 	 * by the first irq.
2289 	 */
2290 	netif_stop_queue(dev);
2291 
2292 
2293 	if (ap->promisc) {
2294 		cmd.evt = C_SET_PROMISC_MODE;
2295 		cmd.code = C_C_PROMISC_DISABLE;
2296 		cmd.idx = 0;
2297 		ace_issue_cmd(regs, &cmd);
2298 		ap->promisc = 0;
2299 	}
2300 
2301 	cmd.evt = C_HOST_STATE;
2302 	cmd.code = C_C_STACK_DOWN;
2303 	cmd.idx = 0;
2304 	ace_issue_cmd(regs, &cmd);
2305 
2306 	tasklet_kill(&ap->ace_tasklet);
2307 
2308 	/*
2309 	 * Make sure one CPU is not processing packets while
2310 	 * buffers are being released by another.
2311 	 */
2312 
2313 	local_irq_save(flags);
2314 	ace_mask_irq(dev);
2315 
2316 	for (i = 0; i < ACE_TX_RING_ENTRIES(ap); i++) {
2317 		struct sk_buff *skb;
2318 		struct tx_ring_info *info;
2319 
2320 		info = ap->skb->tx_skbuff + i;
2321 		skb = info->skb;
2322 
2323 		if (dma_unmap_len(info, maplen)) {
2324 			if (ACE_IS_TIGON_I(ap)) {
2325 				/* NB: TIGON_1 is special, tx_ring is in io space */
2326 				struct tx_desc __iomem *tx;
2327 				tx = (__force struct tx_desc __iomem *) &ap->tx_ring[i];
2328 				writel(0, &tx->addr.addrhi);
2329 				writel(0, &tx->addr.addrlo);
2330 				writel(0, &tx->flagsize);
2331 			} else
2332 				memset(ap->tx_ring + i, 0,
2333 				       sizeof(struct tx_desc));
2334 			dma_unmap_page(&ap->pdev->dev,
2335 				       dma_unmap_addr(info, mapping),
2336 				       dma_unmap_len(info, maplen),
2337 				       DMA_TO_DEVICE);
2338 			dma_unmap_len_set(info, maplen, 0);
2339 		}
2340 		if (skb) {
2341 			dev_kfree_skb(skb);
2342 			info->skb = NULL;
2343 		}
2344 	}
2345 
2346 	if (ap->jumbo) {
2347 		cmd.evt = C_RESET_JUMBO_RNG;
2348 		cmd.code = 0;
2349 		cmd.idx = 0;
2350 		ace_issue_cmd(regs, &cmd);
2351 	}
2352 
2353 	ace_unmask_irq(dev);
2354 	local_irq_restore(flags);
2355 
2356 	return 0;
2357 }
2358 
2359 
2360 static inline dma_addr_t
2361 ace_map_tx_skb(struct ace_private *ap, struct sk_buff *skb,
2362 	       struct sk_buff *tail, u32 idx)
2363 {
2364 	dma_addr_t mapping;
2365 	struct tx_ring_info *info;
2366 
2367 	mapping = dma_map_page(&ap->pdev->dev, virt_to_page(skb->data),
2368 			       offset_in_page(skb->data), skb->len,
2369 			       DMA_TO_DEVICE);
2370 
2371 	info = ap->skb->tx_skbuff + idx;
2372 	info->skb = tail;
2373 	dma_unmap_addr_set(info, mapping, mapping);
2374 	dma_unmap_len_set(info, maplen, skb->len);
2375 	return mapping;
2376 }
2377 
2378 
2379 static inline void
2380 ace_load_tx_bd(struct ace_private *ap, struct tx_desc *desc, u64 addr,
2381 	       u32 flagsize, u32 vlan_tag)
2382 {
2383 #if !USE_TX_COAL_NOW
2384 	flagsize &= ~BD_FLG_COAL_NOW;
2385 #endif
2386 
2387 	if (ACE_IS_TIGON_I(ap)) {
2388 		struct tx_desc __iomem *io = (__force struct tx_desc __iomem *) desc;
2389 		writel(addr >> 32, &io->addr.addrhi);
2390 		writel(addr & 0xffffffff, &io->addr.addrlo);
2391 		writel(flagsize, &io->flagsize);
2392 		writel(vlan_tag, &io->vlanres);
2393 	} else {
2394 		desc->addr.addrhi = addr >> 32;
2395 		desc->addr.addrlo = addr;
2396 		desc->flagsize = flagsize;
2397 		desc->vlanres = vlan_tag;
2398 	}
2399 }
2400 
2401 
2402 static netdev_tx_t ace_start_xmit(struct sk_buff *skb,
2403 				  struct net_device *dev)
2404 {
2405 	struct ace_private *ap = netdev_priv(dev);
2406 	struct ace_regs __iomem *regs = ap->regs;
2407 	struct tx_desc *desc;
2408 	u32 idx, flagsize;
2409 	unsigned long maxjiff = jiffies + 3*HZ;
2410 
2411 restart:
2412 	idx = ap->tx_prd;
2413 
2414 	if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2415 		goto overflow;
2416 
2417 	if (!skb_shinfo(skb)->nr_frags)	{
2418 		dma_addr_t mapping;
2419 		u32 vlan_tag = 0;
2420 
2421 		mapping = ace_map_tx_skb(ap, skb, skb, idx);
2422 		flagsize = (skb->len << 16) | (BD_FLG_END);
2423 		if (skb->ip_summed == CHECKSUM_PARTIAL)
2424 			flagsize |= BD_FLG_TCP_UDP_SUM;
2425 		if (skb_vlan_tag_present(skb)) {
2426 			flagsize |= BD_FLG_VLAN_TAG;
2427 			vlan_tag = skb_vlan_tag_get(skb);
2428 		}
2429 		desc = ap->tx_ring + idx;
2430 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2431 
2432 		/* Look at ace_tx_int for explanations. */
2433 		if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2434 			flagsize |= BD_FLG_COAL_NOW;
2435 
2436 		ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2437 	} else {
2438 		dma_addr_t mapping;
2439 		u32 vlan_tag = 0;
2440 		int i, len = 0;
2441 
2442 		mapping = ace_map_tx_skb(ap, skb, NULL, idx);
2443 		flagsize = (skb_headlen(skb) << 16);
2444 		if (skb->ip_summed == CHECKSUM_PARTIAL)
2445 			flagsize |= BD_FLG_TCP_UDP_SUM;
2446 		if (skb_vlan_tag_present(skb)) {
2447 			flagsize |= BD_FLG_VLAN_TAG;
2448 			vlan_tag = skb_vlan_tag_get(skb);
2449 		}
2450 
2451 		ace_load_tx_bd(ap, ap->tx_ring + idx, mapping, flagsize, vlan_tag);
2452 
2453 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2454 
2455 		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2456 			const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2457 			struct tx_ring_info *info;
2458 
2459 			len += skb_frag_size(frag);
2460 			info = ap->skb->tx_skbuff + idx;
2461 			desc = ap->tx_ring + idx;
2462 
2463 			mapping = skb_frag_dma_map(&ap->pdev->dev, frag, 0,
2464 						   skb_frag_size(frag),
2465 						   DMA_TO_DEVICE);
2466 
2467 			flagsize = skb_frag_size(frag) << 16;
2468 			if (skb->ip_summed == CHECKSUM_PARTIAL)
2469 				flagsize |= BD_FLG_TCP_UDP_SUM;
2470 			idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2471 
2472 			if (i == skb_shinfo(skb)->nr_frags - 1) {
2473 				flagsize |= BD_FLG_END;
2474 				if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2475 					flagsize |= BD_FLG_COAL_NOW;
2476 
2477 				/*
2478 				 * Only the last fragment frees
2479 				 * the skb!
2480 				 */
2481 				info->skb = skb;
2482 			} else {
2483 				info->skb = NULL;
2484 			}
2485 			dma_unmap_addr_set(info, mapping, mapping);
2486 			dma_unmap_len_set(info, maplen, skb_frag_size(frag));
2487 			ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2488 		}
2489 	}
2490 
2491  	wmb();
2492  	ap->tx_prd = idx;
2493  	ace_set_txprd(regs, ap, idx);
2494 
2495 	if (flagsize & BD_FLG_COAL_NOW) {
2496 		netif_stop_queue(dev);
2497 
2498 		/*
2499 		 * A TX-descriptor producer (an IRQ) might have gotten
2500 		 * between, making the ring free again. Since xmit is
2501 		 * serialized, this is the only situation we have to
2502 		 * re-test.
2503 		 */
2504 		if (!tx_ring_full(ap, ap->tx_ret_csm, idx))
2505 			netif_wake_queue(dev);
2506 	}
2507 
2508 	return NETDEV_TX_OK;
2509 
2510 overflow:
2511 	/*
2512 	 * This race condition is unavoidable with lock-free drivers.
2513 	 * We wake up the queue _before_ tx_prd is advanced, so that we can
2514 	 * enter hard_start_xmit too early, while tx ring still looks closed.
2515 	 * This happens ~1-4 times per 100000 packets, so that we can allow
2516 	 * to loop syncing to other CPU. Probably, we need an additional
2517 	 * wmb() in ace_tx_intr as well.
2518 	 *
2519 	 * Note that this race is relieved by reserving one more entry
2520 	 * in tx ring than it is necessary (see original non-SG driver).
2521 	 * However, with SG we need to reserve 2*MAX_SKB_FRAGS+1, which
2522 	 * is already overkill.
2523 	 *
2524 	 * Alternative is to return with 1 not throttling queue. In this
2525 	 * case loop becomes longer, no more useful effects.
2526 	 */
2527 	if (time_before(jiffies, maxjiff)) {
2528 		barrier();
2529 		cpu_relax();
2530 		goto restart;
2531 	}
2532 
2533 	/* The ring is stuck full. */
2534 	printk(KERN_WARNING "%s: Transmit ring stuck full\n", dev->name);
2535 	return NETDEV_TX_BUSY;
2536 }
2537 
2538 
2539 static int ace_change_mtu(struct net_device *dev, int new_mtu)
2540 {
2541 	struct ace_private *ap = netdev_priv(dev);
2542 	struct ace_regs __iomem *regs = ap->regs;
2543 
2544 	writel(new_mtu + ETH_HLEN + 4, &regs->IfMtu);
2545 	dev->mtu = new_mtu;
2546 
2547 	if (new_mtu > ACE_STD_MTU) {
2548 		if (!(ap->jumbo)) {
2549 			printk(KERN_INFO "%s: Enabling Jumbo frame "
2550 			       "support\n", dev->name);
2551 			ap->jumbo = 1;
2552 			if (!test_and_set_bit(0, &ap->jumbo_refill_busy))
2553 				ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE);
2554 			ace_set_rxtx_parms(dev, 1);
2555 		}
2556 	} else {
2557 		while (test_and_set_bit(0, &ap->jumbo_refill_busy));
2558 		ace_sync_irq(dev->irq);
2559 		ace_set_rxtx_parms(dev, 0);
2560 		if (ap->jumbo) {
2561 			struct cmd cmd;
2562 
2563 			cmd.evt = C_RESET_JUMBO_RNG;
2564 			cmd.code = 0;
2565 			cmd.idx = 0;
2566 			ace_issue_cmd(regs, &cmd);
2567 		}
2568 	}
2569 
2570 	return 0;
2571 }
2572 
2573 static int ace_get_link_ksettings(struct net_device *dev,
2574 				  struct ethtool_link_ksettings *cmd)
2575 {
2576 	struct ace_private *ap = netdev_priv(dev);
2577 	struct ace_regs __iomem *regs = ap->regs;
2578 	u32 link;
2579 	u32 supported;
2580 
2581 	memset(cmd, 0, sizeof(struct ethtool_link_ksettings));
2582 
2583 	supported = (SUPPORTED_10baseT_Half | SUPPORTED_10baseT_Full |
2584 		     SUPPORTED_100baseT_Half | SUPPORTED_100baseT_Full |
2585 		     SUPPORTED_1000baseT_Half | SUPPORTED_1000baseT_Full |
2586 		     SUPPORTED_Autoneg | SUPPORTED_FIBRE);
2587 
2588 	cmd->base.port = PORT_FIBRE;
2589 
2590 	link = readl(&regs->GigLnkState);
2591 	if (link & LNK_1000MB) {
2592 		cmd->base.speed = SPEED_1000;
2593 	} else {
2594 		link = readl(&regs->FastLnkState);
2595 		if (link & LNK_100MB)
2596 			cmd->base.speed = SPEED_100;
2597 		else if (link & LNK_10MB)
2598 			cmd->base.speed = SPEED_10;
2599 		else
2600 			cmd->base.speed = 0;
2601 	}
2602 	if (link & LNK_FULL_DUPLEX)
2603 		cmd->base.duplex = DUPLEX_FULL;
2604 	else
2605 		cmd->base.duplex = DUPLEX_HALF;
2606 
2607 	if (link & LNK_NEGOTIATE)
2608 		cmd->base.autoneg = AUTONEG_ENABLE;
2609 	else
2610 		cmd->base.autoneg = AUTONEG_DISABLE;
2611 
2612 #if 0
2613 	/*
2614 	 * Current struct ethtool_cmd is insufficient
2615 	 */
2616 	ecmd->trace = readl(&regs->TuneTrace);
2617 
2618 	ecmd->txcoal = readl(&regs->TuneTxCoalTicks);
2619 	ecmd->rxcoal = readl(&regs->TuneRxCoalTicks);
2620 #endif
2621 
2622 	ethtool_convert_legacy_u32_to_link_mode(cmd->link_modes.supported,
2623 						supported);
2624 
2625 	return 0;
2626 }
2627 
2628 static int ace_set_link_ksettings(struct net_device *dev,
2629 				  const struct ethtool_link_ksettings *cmd)
2630 {
2631 	struct ace_private *ap = netdev_priv(dev);
2632 	struct ace_regs __iomem *regs = ap->regs;
2633 	u32 link, speed;
2634 
2635 	link = readl(&regs->GigLnkState);
2636 	if (link & LNK_1000MB)
2637 		speed = SPEED_1000;
2638 	else {
2639 		link = readl(&regs->FastLnkState);
2640 		if (link & LNK_100MB)
2641 			speed = SPEED_100;
2642 		else if (link & LNK_10MB)
2643 			speed = SPEED_10;
2644 		else
2645 			speed = SPEED_100;
2646 	}
2647 
2648 	link = LNK_ENABLE | LNK_1000MB | LNK_100MB | LNK_10MB |
2649 		LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL;
2650 	if (!ACE_IS_TIGON_I(ap))
2651 		link |= LNK_TX_FLOW_CTL_Y;
2652 	if (cmd->base.autoneg == AUTONEG_ENABLE)
2653 		link |= LNK_NEGOTIATE;
2654 	if (cmd->base.speed != speed) {
2655 		link &= ~(LNK_1000MB | LNK_100MB | LNK_10MB);
2656 		switch (cmd->base.speed) {
2657 		case SPEED_1000:
2658 			link |= LNK_1000MB;
2659 			break;
2660 		case SPEED_100:
2661 			link |= LNK_100MB;
2662 			break;
2663 		case SPEED_10:
2664 			link |= LNK_10MB;
2665 			break;
2666 		}
2667 	}
2668 
2669 	if (cmd->base.duplex == DUPLEX_FULL)
2670 		link |= LNK_FULL_DUPLEX;
2671 
2672 	if (link != ap->link) {
2673 		struct cmd cmd;
2674 		printk(KERN_INFO "%s: Renegotiating link state\n",
2675 		       dev->name);
2676 
2677 		ap->link = link;
2678 		writel(link, &regs->TuneLink);
2679 		if (!ACE_IS_TIGON_I(ap))
2680 			writel(link, &regs->TuneFastLink);
2681 		wmb();
2682 
2683 		cmd.evt = C_LNK_NEGOTIATION;
2684 		cmd.code = 0;
2685 		cmd.idx = 0;
2686 		ace_issue_cmd(regs, &cmd);
2687 	}
2688 	return 0;
2689 }
2690 
2691 static void ace_get_drvinfo(struct net_device *dev,
2692 			    struct ethtool_drvinfo *info)
2693 {
2694 	struct ace_private *ap = netdev_priv(dev);
2695 
2696 	strlcpy(info->driver, "acenic", sizeof(info->driver));
2697 	snprintf(info->fw_version, sizeof(info->version), "%i.%i.%i",
2698 		 ap->firmware_major, ap->firmware_minor, ap->firmware_fix);
2699 
2700 	if (ap->pdev)
2701 		strlcpy(info->bus_info, pci_name(ap->pdev),
2702 			sizeof(info->bus_info));
2703 
2704 }
2705 
2706 /*
2707  * Set the hardware MAC address.
2708  */
2709 static int ace_set_mac_addr(struct net_device *dev, void *p)
2710 {
2711 	struct ace_private *ap = netdev_priv(dev);
2712 	struct ace_regs __iomem *regs = ap->regs;
2713 	struct sockaddr *addr=p;
2714 	u8 *da;
2715 	struct cmd cmd;
2716 
2717 	if(netif_running(dev))
2718 		return -EBUSY;
2719 
2720 	memcpy(dev->dev_addr, addr->sa_data,dev->addr_len);
2721 
2722 	da = (u8 *)dev->dev_addr;
2723 
2724 	writel(da[0] << 8 | da[1], &regs->MacAddrHi);
2725 	writel((da[2] << 24) | (da[3] << 16) | (da[4] << 8) | da[5],
2726 	       &regs->MacAddrLo);
2727 
2728 	cmd.evt = C_SET_MAC_ADDR;
2729 	cmd.code = 0;
2730 	cmd.idx = 0;
2731 	ace_issue_cmd(regs, &cmd);
2732 
2733 	return 0;
2734 }
2735 
2736 
2737 static void ace_set_multicast_list(struct net_device *dev)
2738 {
2739 	struct ace_private *ap = netdev_priv(dev);
2740 	struct ace_regs __iomem *regs = ap->regs;
2741 	struct cmd cmd;
2742 
2743 	if ((dev->flags & IFF_ALLMULTI) && !(ap->mcast_all)) {
2744 		cmd.evt = C_SET_MULTICAST_MODE;
2745 		cmd.code = C_C_MCAST_ENABLE;
2746 		cmd.idx = 0;
2747 		ace_issue_cmd(regs, &cmd);
2748 		ap->mcast_all = 1;
2749 	} else if (ap->mcast_all) {
2750 		cmd.evt = C_SET_MULTICAST_MODE;
2751 		cmd.code = C_C_MCAST_DISABLE;
2752 		cmd.idx = 0;
2753 		ace_issue_cmd(regs, &cmd);
2754 		ap->mcast_all = 0;
2755 	}
2756 
2757 	if ((dev->flags & IFF_PROMISC) && !(ap->promisc)) {
2758 		cmd.evt = C_SET_PROMISC_MODE;
2759 		cmd.code = C_C_PROMISC_ENABLE;
2760 		cmd.idx = 0;
2761 		ace_issue_cmd(regs, &cmd);
2762 		ap->promisc = 1;
2763 	}else if (!(dev->flags & IFF_PROMISC) && (ap->promisc)) {
2764 		cmd.evt = C_SET_PROMISC_MODE;
2765 		cmd.code = C_C_PROMISC_DISABLE;
2766 		cmd.idx = 0;
2767 		ace_issue_cmd(regs, &cmd);
2768 		ap->promisc = 0;
2769 	}
2770 
2771 	/*
2772 	 * For the time being multicast relies on the upper layers
2773 	 * filtering it properly. The Firmware does not allow one to
2774 	 * set the entire multicast list at a time and keeping track of
2775 	 * it here is going to be messy.
2776 	 */
2777 	if (!netdev_mc_empty(dev) && !ap->mcast_all) {
2778 		cmd.evt = C_SET_MULTICAST_MODE;
2779 		cmd.code = C_C_MCAST_ENABLE;
2780 		cmd.idx = 0;
2781 		ace_issue_cmd(regs, &cmd);
2782 	}else if (!ap->mcast_all) {
2783 		cmd.evt = C_SET_MULTICAST_MODE;
2784 		cmd.code = C_C_MCAST_DISABLE;
2785 		cmd.idx = 0;
2786 		ace_issue_cmd(regs, &cmd);
2787 	}
2788 }
2789 
2790 
2791 static struct net_device_stats *ace_get_stats(struct net_device *dev)
2792 {
2793 	struct ace_private *ap = netdev_priv(dev);
2794 	struct ace_mac_stats __iomem *mac_stats =
2795 		(struct ace_mac_stats __iomem *)ap->regs->Stats;
2796 
2797 	dev->stats.rx_missed_errors = readl(&mac_stats->drop_space);
2798 	dev->stats.multicast = readl(&mac_stats->kept_mc);
2799 	dev->stats.collisions = readl(&mac_stats->coll);
2800 
2801 	return &dev->stats;
2802 }
2803 
2804 
2805 static void ace_copy(struct ace_regs __iomem *regs, const __be32 *src,
2806 		     u32 dest, int size)
2807 {
2808 	void __iomem *tdest;
2809 	short tsize, i;
2810 
2811 	if (size <= 0)
2812 		return;
2813 
2814 	while (size > 0) {
2815 		tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2816 			    min_t(u32, size, ACE_WINDOW_SIZE));
2817 		tdest = (void __iomem *) &regs->Window +
2818 			(dest & (ACE_WINDOW_SIZE - 1));
2819 		writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2820 		for (i = 0; i < (tsize / 4); i++) {
2821 			/* Firmware is big-endian */
2822 			writel(be32_to_cpup(src), tdest);
2823 			src++;
2824 			tdest += 4;
2825 			dest += 4;
2826 			size -= 4;
2827 		}
2828 	}
2829 }
2830 
2831 
2832 static void ace_clear(struct ace_regs __iomem *regs, u32 dest, int size)
2833 {
2834 	void __iomem *tdest;
2835 	short tsize = 0, i;
2836 
2837 	if (size <= 0)
2838 		return;
2839 
2840 	while (size > 0) {
2841 		tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2842 				min_t(u32, size, ACE_WINDOW_SIZE));
2843 		tdest = (void __iomem *) &regs->Window +
2844 			(dest & (ACE_WINDOW_SIZE - 1));
2845 		writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2846 
2847 		for (i = 0; i < (tsize / 4); i++) {
2848 			writel(0, tdest + i*4);
2849 		}
2850 
2851 		dest += tsize;
2852 		size -= tsize;
2853 	}
2854 }
2855 
2856 
2857 /*
2858  * Download the firmware into the SRAM on the NIC
2859  *
2860  * This operation requires the NIC to be halted and is performed with
2861  * interrupts disabled and with the spinlock hold.
2862  */
2863 static int ace_load_firmware(struct net_device *dev)
2864 {
2865 	const struct firmware *fw;
2866 	const char *fw_name = "acenic/tg2.bin";
2867 	struct ace_private *ap = netdev_priv(dev);
2868 	struct ace_regs __iomem *regs = ap->regs;
2869 	const __be32 *fw_data;
2870 	u32 load_addr;
2871 	int ret;
2872 
2873 	if (!(readl(&regs->CpuCtrl) & CPU_HALTED)) {
2874 		printk(KERN_ERR "%s: trying to download firmware while the "
2875 		       "CPU is running!\n", ap->name);
2876 		return -EFAULT;
2877 	}
2878 
2879 	if (ACE_IS_TIGON_I(ap))
2880 		fw_name = "acenic/tg1.bin";
2881 
2882 	ret = request_firmware(&fw, fw_name, &ap->pdev->dev);
2883 	if (ret) {
2884 		printk(KERN_ERR "%s: Failed to load firmware \"%s\"\n",
2885 		       ap->name, fw_name);
2886 		return ret;
2887 	}
2888 
2889 	fw_data = (void *)fw->data;
2890 
2891 	/* Firmware blob starts with version numbers, followed by
2892 	   load and start address. Remainder is the blob to be loaded
2893 	   contiguously from load address. We don't bother to represent
2894 	   the BSS/SBSS sections any more, since we were clearing the
2895 	   whole thing anyway. */
2896 	ap->firmware_major = fw->data[0];
2897 	ap->firmware_minor = fw->data[1];
2898 	ap->firmware_fix = fw->data[2];
2899 
2900 	ap->firmware_start = be32_to_cpu(fw_data[1]);
2901 	if (ap->firmware_start < 0x4000 || ap->firmware_start >= 0x80000) {
2902 		printk(KERN_ERR "%s: bogus load address %08x in \"%s\"\n",
2903 		       ap->name, ap->firmware_start, fw_name);
2904 		ret = -EINVAL;
2905 		goto out;
2906 	}
2907 
2908 	load_addr = be32_to_cpu(fw_data[2]);
2909 	if (load_addr < 0x4000 || load_addr >= 0x80000) {
2910 		printk(KERN_ERR "%s: bogus load address %08x in \"%s\"\n",
2911 		       ap->name, load_addr, fw_name);
2912 		ret = -EINVAL;
2913 		goto out;
2914 	}
2915 
2916 	/*
2917 	 * Do not try to clear more than 512KiB or we end up seeing
2918 	 * funny things on NICs with only 512KiB SRAM
2919 	 */
2920 	ace_clear(regs, 0x2000, 0x80000-0x2000);
2921 	ace_copy(regs, &fw_data[3], load_addr, fw->size-12);
2922  out:
2923 	release_firmware(fw);
2924 	return ret;
2925 }
2926 
2927 
2928 /*
2929  * The eeprom on the AceNIC is an Atmel i2c EEPROM.
2930  *
2931  * Accessing the EEPROM is `interesting' to say the least - don't read
2932  * this code right after dinner.
2933  *
2934  * This is all about black magic and bit-banging the device .... I
2935  * wonder in what hospital they have put the guy who designed the i2c
2936  * specs.
2937  *
2938  * Oh yes, this is only the beginning!
2939  *
2940  * Thanks to Stevarino Webinski for helping tracking down the bugs in the
2941  * code i2c readout code by beta testing all my hacks.
2942  */
2943 static void eeprom_start(struct ace_regs __iomem *regs)
2944 {
2945 	u32 local;
2946 
2947 	readl(&regs->LocalCtrl);
2948 	udelay(ACE_SHORT_DELAY);
2949 	local = readl(&regs->LocalCtrl);
2950 	local |= EEPROM_DATA_OUT | EEPROM_WRITE_ENABLE;
2951 	writel(local, &regs->LocalCtrl);
2952 	readl(&regs->LocalCtrl);
2953 	mb();
2954 	udelay(ACE_SHORT_DELAY);
2955 	local |= EEPROM_CLK_OUT;
2956 	writel(local, &regs->LocalCtrl);
2957 	readl(&regs->LocalCtrl);
2958 	mb();
2959 	udelay(ACE_SHORT_DELAY);
2960 	local &= ~EEPROM_DATA_OUT;
2961 	writel(local, &regs->LocalCtrl);
2962 	readl(&regs->LocalCtrl);
2963 	mb();
2964 	udelay(ACE_SHORT_DELAY);
2965 	local &= ~EEPROM_CLK_OUT;
2966 	writel(local, &regs->LocalCtrl);
2967 	readl(&regs->LocalCtrl);
2968 	mb();
2969 }
2970 
2971 
2972 static void eeprom_prep(struct ace_regs __iomem *regs, u8 magic)
2973 {
2974 	short i;
2975 	u32 local;
2976 
2977 	udelay(ACE_SHORT_DELAY);
2978 	local = readl(&regs->LocalCtrl);
2979 	local &= ~EEPROM_DATA_OUT;
2980 	local |= EEPROM_WRITE_ENABLE;
2981 	writel(local, &regs->LocalCtrl);
2982 	readl(&regs->LocalCtrl);
2983 	mb();
2984 
2985 	for (i = 0; i < 8; i++, magic <<= 1) {
2986 		udelay(ACE_SHORT_DELAY);
2987 		if (magic & 0x80)
2988 			local |= EEPROM_DATA_OUT;
2989 		else
2990 			local &= ~EEPROM_DATA_OUT;
2991 		writel(local, &regs->LocalCtrl);
2992 		readl(&regs->LocalCtrl);
2993 		mb();
2994 
2995 		udelay(ACE_SHORT_DELAY);
2996 		local |= EEPROM_CLK_OUT;
2997 		writel(local, &regs->LocalCtrl);
2998 		readl(&regs->LocalCtrl);
2999 		mb();
3000 		udelay(ACE_SHORT_DELAY);
3001 		local &= ~(EEPROM_CLK_OUT | EEPROM_DATA_OUT);
3002 		writel(local, &regs->LocalCtrl);
3003 		readl(&regs->LocalCtrl);
3004 		mb();
3005 	}
3006 }
3007 
3008 
3009 static int eeprom_check_ack(struct ace_regs __iomem *regs)
3010 {
3011 	int state;
3012 	u32 local;
3013 
3014 	local = readl(&regs->LocalCtrl);
3015 	local &= ~EEPROM_WRITE_ENABLE;
3016 	writel(local, &regs->LocalCtrl);
3017 	readl(&regs->LocalCtrl);
3018 	mb();
3019 	udelay(ACE_LONG_DELAY);
3020 	local |= EEPROM_CLK_OUT;
3021 	writel(local, &regs->LocalCtrl);
3022 	readl(&regs->LocalCtrl);
3023 	mb();
3024 	udelay(ACE_SHORT_DELAY);
3025 	/* sample data in middle of high clk */
3026 	state = (readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0;
3027 	udelay(ACE_SHORT_DELAY);
3028 	mb();
3029 	writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3030 	readl(&regs->LocalCtrl);
3031 	mb();
3032 
3033 	return state;
3034 }
3035 
3036 
3037 static void eeprom_stop(struct ace_regs __iomem *regs)
3038 {
3039 	u32 local;
3040 
3041 	udelay(ACE_SHORT_DELAY);
3042 	local = readl(&regs->LocalCtrl);
3043 	local |= EEPROM_WRITE_ENABLE;
3044 	writel(local, &regs->LocalCtrl);
3045 	readl(&regs->LocalCtrl);
3046 	mb();
3047 	udelay(ACE_SHORT_DELAY);
3048 	local &= ~EEPROM_DATA_OUT;
3049 	writel(local, &regs->LocalCtrl);
3050 	readl(&regs->LocalCtrl);
3051 	mb();
3052 	udelay(ACE_SHORT_DELAY);
3053 	local |= EEPROM_CLK_OUT;
3054 	writel(local, &regs->LocalCtrl);
3055 	readl(&regs->LocalCtrl);
3056 	mb();
3057 	udelay(ACE_SHORT_DELAY);
3058 	local |= EEPROM_DATA_OUT;
3059 	writel(local, &regs->LocalCtrl);
3060 	readl(&regs->LocalCtrl);
3061 	mb();
3062 	udelay(ACE_LONG_DELAY);
3063 	local &= ~EEPROM_CLK_OUT;
3064 	writel(local, &regs->LocalCtrl);
3065 	mb();
3066 }
3067 
3068 
3069 /*
3070  * Read a whole byte from the EEPROM.
3071  */
3072 static int read_eeprom_byte(struct net_device *dev, unsigned long offset)
3073 {
3074 	struct ace_private *ap = netdev_priv(dev);
3075 	struct ace_regs __iomem *regs = ap->regs;
3076 	unsigned long flags;
3077 	u32 local;
3078 	int result = 0;
3079 	short i;
3080 
3081 	/*
3082 	 * Don't take interrupts on this CPU will bit banging
3083 	 * the %#%#@$ I2C device
3084 	 */
3085 	local_irq_save(flags);
3086 
3087 	eeprom_start(regs);
3088 
3089 	eeprom_prep(regs, EEPROM_WRITE_SELECT);
3090 	if (eeprom_check_ack(regs)) {
3091 		local_irq_restore(flags);
3092 		printk(KERN_ERR "%s: Unable to sync eeprom\n", ap->name);
3093 		result = -EIO;
3094 		goto eeprom_read_error;
3095 	}
3096 
3097 	eeprom_prep(regs, (offset >> 8) & 0xff);
3098 	if (eeprom_check_ack(regs)) {
3099 		local_irq_restore(flags);
3100 		printk(KERN_ERR "%s: Unable to set address byte 0\n",
3101 		       ap->name);
3102 		result = -EIO;
3103 		goto eeprom_read_error;
3104 	}
3105 
3106 	eeprom_prep(regs, offset & 0xff);
3107 	if (eeprom_check_ack(regs)) {
3108 		local_irq_restore(flags);
3109 		printk(KERN_ERR "%s: Unable to set address byte 1\n",
3110 		       ap->name);
3111 		result = -EIO;
3112 		goto eeprom_read_error;
3113 	}
3114 
3115 	eeprom_start(regs);
3116 	eeprom_prep(regs, EEPROM_READ_SELECT);
3117 	if (eeprom_check_ack(regs)) {
3118 		local_irq_restore(flags);
3119 		printk(KERN_ERR "%s: Unable to set READ_SELECT\n",
3120 		       ap->name);
3121 		result = -EIO;
3122 		goto eeprom_read_error;
3123 	}
3124 
3125 	for (i = 0; i < 8; i++) {
3126 		local = readl(&regs->LocalCtrl);
3127 		local &= ~EEPROM_WRITE_ENABLE;
3128 		writel(local, &regs->LocalCtrl);
3129 		readl(&regs->LocalCtrl);
3130 		udelay(ACE_LONG_DELAY);
3131 		mb();
3132 		local |= EEPROM_CLK_OUT;
3133 		writel(local, &regs->LocalCtrl);
3134 		readl(&regs->LocalCtrl);
3135 		mb();
3136 		udelay(ACE_SHORT_DELAY);
3137 		/* sample data mid high clk */
3138 		result = (result << 1) |
3139 			((readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0);
3140 		udelay(ACE_SHORT_DELAY);
3141 		mb();
3142 		local = readl(&regs->LocalCtrl);
3143 		local &= ~EEPROM_CLK_OUT;
3144 		writel(local, &regs->LocalCtrl);
3145 		readl(&regs->LocalCtrl);
3146 		udelay(ACE_SHORT_DELAY);
3147 		mb();
3148 		if (i == 7) {
3149 			local |= EEPROM_WRITE_ENABLE;
3150 			writel(local, &regs->LocalCtrl);
3151 			readl(&regs->LocalCtrl);
3152 			mb();
3153 			udelay(ACE_SHORT_DELAY);
3154 		}
3155 	}
3156 
3157 	local |= EEPROM_DATA_OUT;
3158 	writel(local, &regs->LocalCtrl);
3159 	readl(&regs->LocalCtrl);
3160 	mb();
3161 	udelay(ACE_SHORT_DELAY);
3162 	writel(readl(&regs->LocalCtrl) | EEPROM_CLK_OUT, &regs->LocalCtrl);
3163 	readl(&regs->LocalCtrl);
3164 	udelay(ACE_LONG_DELAY);
3165 	writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3166 	readl(&regs->LocalCtrl);
3167 	mb();
3168 	udelay(ACE_SHORT_DELAY);
3169 	eeprom_stop(regs);
3170 
3171 	local_irq_restore(flags);
3172  out:
3173 	return result;
3174 
3175  eeprom_read_error:
3176 	printk(KERN_ERR "%s: Unable to read eeprom byte 0x%02lx\n",
3177 	       ap->name, offset);
3178 	goto out;
3179 }
3180 
3181 module_pci_driver(acenic_pci_driver);
3182