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->ndev = dev;
469 	ap->pdev = pdev;
470 	ap->name = pci_name(pdev);
471 
472 	dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
473 	dev->features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX;
474 
475 	dev->watchdog_timeo = 5*HZ;
476 	dev->min_mtu = 0;
477 	dev->max_mtu = ACE_JUMBO_MTU;
478 
479 	dev->netdev_ops = &ace_netdev_ops;
480 	dev->ethtool_ops = &ace_ethtool_ops;
481 
482 	/* we only display this string ONCE */
483 	if (!boards_found)
484 		printk(version);
485 
486 	if (pci_enable_device(pdev))
487 		goto fail_free_netdev;
488 
489 	/*
490 	 * Enable master mode before we start playing with the
491 	 * pci_command word since pci_set_master() will modify
492 	 * it.
493 	 */
494 	pci_set_master(pdev);
495 
496 	pci_read_config_word(pdev, PCI_COMMAND, &ap->pci_command);
497 
498 	/* OpenFirmware on Mac's does not set this - DOH.. */
499 	if (!(ap->pci_command & PCI_COMMAND_MEMORY)) {
500 		printk(KERN_INFO "%s: Enabling PCI Memory Mapped "
501 		       "access - was not enabled by BIOS/Firmware\n",
502 		       ap->name);
503 		ap->pci_command = ap->pci_command | PCI_COMMAND_MEMORY;
504 		pci_write_config_word(ap->pdev, PCI_COMMAND,
505 				      ap->pci_command);
506 		wmb();
507 	}
508 
509 	pci_read_config_byte(pdev, PCI_LATENCY_TIMER, &ap->pci_latency);
510 	if (ap->pci_latency <= 0x40) {
511 		ap->pci_latency = 0x40;
512 		pci_write_config_byte(pdev, PCI_LATENCY_TIMER, ap->pci_latency);
513 	}
514 
515 	/*
516 	 * Remap the regs into kernel space - this is abuse of
517 	 * dev->base_addr since it was means for I/O port
518 	 * addresses but who gives a damn.
519 	 */
520 	dev->base_addr = pci_resource_start(pdev, 0);
521 	ap->regs = ioremap(dev->base_addr, 0x4000);
522 	if (!ap->regs) {
523 		printk(KERN_ERR "%s:  Unable to map I/O register, "
524 		       "AceNIC %i will be disabled.\n",
525 		       ap->name, boards_found);
526 		goto fail_free_netdev;
527 	}
528 
529 	switch(pdev->vendor) {
530 	case PCI_VENDOR_ID_ALTEON:
531 		if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9100T) {
532 			printk(KERN_INFO "%s: Farallon PN9100-T ",
533 			       ap->name);
534 		} else {
535 			printk(KERN_INFO "%s: Alteon AceNIC ",
536 			       ap->name);
537 		}
538 		break;
539 	case PCI_VENDOR_ID_3COM:
540 		printk(KERN_INFO "%s: 3Com 3C985 ", ap->name);
541 		break;
542 	case PCI_VENDOR_ID_NETGEAR:
543 		printk(KERN_INFO "%s: NetGear GA620 ", ap->name);
544 		break;
545 	case PCI_VENDOR_ID_DEC:
546 		if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9000SX) {
547 			printk(KERN_INFO "%s: Farallon PN9000-SX ",
548 			       ap->name);
549 			break;
550 		}
551 		fallthrough;
552 	case PCI_VENDOR_ID_SGI:
553 		printk(KERN_INFO "%s: SGI AceNIC ", ap->name);
554 		break;
555 	default:
556 		printk(KERN_INFO "%s: Unknown AceNIC ", ap->name);
557 		break;
558 	}
559 
560 	printk("Gigabit Ethernet at 0x%08lx, ", dev->base_addr);
561 	printk("irq %d\n", pdev->irq);
562 
563 #ifdef CONFIG_ACENIC_OMIT_TIGON_I
564 	if ((readl(&ap->regs->HostCtrl) >> 28) == 4) {
565 		printk(KERN_ERR "%s: Driver compiled without Tigon I"
566 		       " support - NIC disabled\n", dev->name);
567 		goto fail_uninit;
568 	}
569 #endif
570 
571 	if (ace_allocate_descriptors(dev))
572 		goto fail_free_netdev;
573 
574 #ifdef MODULE
575 	if (boards_found >= ACE_MAX_MOD_PARMS)
576 		ap->board_idx = BOARD_IDX_OVERFLOW;
577 	else
578 		ap->board_idx = boards_found;
579 #else
580 	ap->board_idx = BOARD_IDX_STATIC;
581 #endif
582 
583 	if (ace_init(dev))
584 		goto fail_free_netdev;
585 
586 	if (register_netdev(dev)) {
587 		printk(KERN_ERR "acenic: device registration failed\n");
588 		goto fail_uninit;
589 	}
590 	ap->name = dev->name;
591 
592 	if (ap->pci_using_dac)
593 		dev->features |= NETIF_F_HIGHDMA;
594 
595 	pci_set_drvdata(pdev, dev);
596 
597 	boards_found++;
598 	return 0;
599 
600  fail_uninit:
601 	ace_init_cleanup(dev);
602  fail_free_netdev:
603 	free_netdev(dev);
604 	return -ENODEV;
605 }
606 
607 static void acenic_remove_one(struct pci_dev *pdev)
608 {
609 	struct net_device *dev = pci_get_drvdata(pdev);
610 	struct ace_private *ap = netdev_priv(dev);
611 	struct ace_regs __iomem *regs = ap->regs;
612 	short i;
613 
614 	unregister_netdev(dev);
615 
616 	writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
617 	if (ap->version >= 2)
618 		writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
619 
620 	/*
621 	 * This clears any pending interrupts
622 	 */
623 	writel(1, &regs->Mb0Lo);
624 	readl(&regs->CpuCtrl);	/* flush */
625 
626 	/*
627 	 * Make sure no other CPUs are processing interrupts
628 	 * on the card before the buffers are being released.
629 	 * Otherwise one might experience some `interesting'
630 	 * effects.
631 	 *
632 	 * Then release the RX buffers - jumbo buffers were
633 	 * already released in ace_close().
634 	 */
635 	ace_sync_irq(dev->irq);
636 
637 	for (i = 0; i < RX_STD_RING_ENTRIES; i++) {
638 		struct sk_buff *skb = ap->skb->rx_std_skbuff[i].skb;
639 
640 		if (skb) {
641 			struct ring_info *ringp;
642 			dma_addr_t mapping;
643 
644 			ringp = &ap->skb->rx_std_skbuff[i];
645 			mapping = dma_unmap_addr(ringp, mapping);
646 			dma_unmap_page(&ap->pdev->dev, mapping,
647 				       ACE_STD_BUFSIZE, DMA_FROM_DEVICE);
648 
649 			ap->rx_std_ring[i].size = 0;
650 			ap->skb->rx_std_skbuff[i].skb = NULL;
651 			dev_kfree_skb(skb);
652 		}
653 	}
654 
655 	if (ap->version >= 2) {
656 		for (i = 0; i < RX_MINI_RING_ENTRIES; i++) {
657 			struct sk_buff *skb = ap->skb->rx_mini_skbuff[i].skb;
658 
659 			if (skb) {
660 				struct ring_info *ringp;
661 				dma_addr_t mapping;
662 
663 				ringp = &ap->skb->rx_mini_skbuff[i];
664 				mapping = dma_unmap_addr(ringp,mapping);
665 				dma_unmap_page(&ap->pdev->dev, mapping,
666 					       ACE_MINI_BUFSIZE,
667 					       DMA_FROM_DEVICE);
668 
669 				ap->rx_mini_ring[i].size = 0;
670 				ap->skb->rx_mini_skbuff[i].skb = NULL;
671 				dev_kfree_skb(skb);
672 			}
673 		}
674 	}
675 
676 	for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
677 		struct sk_buff *skb = ap->skb->rx_jumbo_skbuff[i].skb;
678 		if (skb) {
679 			struct ring_info *ringp;
680 			dma_addr_t mapping;
681 
682 			ringp = &ap->skb->rx_jumbo_skbuff[i];
683 			mapping = dma_unmap_addr(ringp, mapping);
684 			dma_unmap_page(&ap->pdev->dev, mapping,
685 				       ACE_JUMBO_BUFSIZE, DMA_FROM_DEVICE);
686 
687 			ap->rx_jumbo_ring[i].size = 0;
688 			ap->skb->rx_jumbo_skbuff[i].skb = NULL;
689 			dev_kfree_skb(skb);
690 		}
691 	}
692 
693 	ace_init_cleanup(dev);
694 	free_netdev(dev);
695 }
696 
697 static struct pci_driver acenic_pci_driver = {
698 	.name		= "acenic",
699 	.id_table	= acenic_pci_tbl,
700 	.probe		= acenic_probe_one,
701 	.remove		= acenic_remove_one,
702 };
703 
704 static void ace_free_descriptors(struct net_device *dev)
705 {
706 	struct ace_private *ap = netdev_priv(dev);
707 	int size;
708 
709 	if (ap->rx_std_ring != NULL) {
710 		size = (sizeof(struct rx_desc) *
711 			(RX_STD_RING_ENTRIES +
712 			 RX_JUMBO_RING_ENTRIES +
713 			 RX_MINI_RING_ENTRIES +
714 			 RX_RETURN_RING_ENTRIES));
715 		dma_free_coherent(&ap->pdev->dev, size, ap->rx_std_ring,
716 				  ap->rx_ring_base_dma);
717 		ap->rx_std_ring = NULL;
718 		ap->rx_jumbo_ring = NULL;
719 		ap->rx_mini_ring = NULL;
720 		ap->rx_return_ring = NULL;
721 	}
722 	if (ap->evt_ring != NULL) {
723 		size = (sizeof(struct event) * EVT_RING_ENTRIES);
724 		dma_free_coherent(&ap->pdev->dev, size, ap->evt_ring,
725 				  ap->evt_ring_dma);
726 		ap->evt_ring = NULL;
727 	}
728 	if (ap->tx_ring != NULL && !ACE_IS_TIGON_I(ap)) {
729 		size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
730 		dma_free_coherent(&ap->pdev->dev, size, ap->tx_ring,
731 				  ap->tx_ring_dma);
732 	}
733 	ap->tx_ring = NULL;
734 
735 	if (ap->evt_prd != NULL) {
736 		dma_free_coherent(&ap->pdev->dev, sizeof(u32),
737 				  (void *)ap->evt_prd, ap->evt_prd_dma);
738 		ap->evt_prd = NULL;
739 	}
740 	if (ap->rx_ret_prd != NULL) {
741 		dma_free_coherent(&ap->pdev->dev, sizeof(u32),
742 				  (void *)ap->rx_ret_prd, ap->rx_ret_prd_dma);
743 		ap->rx_ret_prd = NULL;
744 	}
745 	if (ap->tx_csm != NULL) {
746 		dma_free_coherent(&ap->pdev->dev, sizeof(u32),
747 				  (void *)ap->tx_csm, ap->tx_csm_dma);
748 		ap->tx_csm = NULL;
749 	}
750 }
751 
752 
753 static int ace_allocate_descriptors(struct net_device *dev)
754 {
755 	struct ace_private *ap = netdev_priv(dev);
756 	int size;
757 
758 	size = (sizeof(struct rx_desc) *
759 		(RX_STD_RING_ENTRIES +
760 		 RX_JUMBO_RING_ENTRIES +
761 		 RX_MINI_RING_ENTRIES +
762 		 RX_RETURN_RING_ENTRIES));
763 
764 	ap->rx_std_ring = dma_alloc_coherent(&ap->pdev->dev, size,
765 					     &ap->rx_ring_base_dma, GFP_KERNEL);
766 	if (ap->rx_std_ring == NULL)
767 		goto fail;
768 
769 	ap->rx_jumbo_ring = ap->rx_std_ring + RX_STD_RING_ENTRIES;
770 	ap->rx_mini_ring = ap->rx_jumbo_ring + RX_JUMBO_RING_ENTRIES;
771 	ap->rx_return_ring = ap->rx_mini_ring + RX_MINI_RING_ENTRIES;
772 
773 	size = (sizeof(struct event) * EVT_RING_ENTRIES);
774 
775 	ap->evt_ring = dma_alloc_coherent(&ap->pdev->dev, size,
776 					  &ap->evt_ring_dma, GFP_KERNEL);
777 
778 	if (ap->evt_ring == NULL)
779 		goto fail;
780 
781 	/*
782 	 * Only allocate a host TX ring for the Tigon II, the Tigon I
783 	 * has to use PCI registers for this ;-(
784 	 */
785 	if (!ACE_IS_TIGON_I(ap)) {
786 		size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
787 
788 		ap->tx_ring = dma_alloc_coherent(&ap->pdev->dev, size,
789 						 &ap->tx_ring_dma, GFP_KERNEL);
790 
791 		if (ap->tx_ring == NULL)
792 			goto fail;
793 	}
794 
795 	ap->evt_prd = dma_alloc_coherent(&ap->pdev->dev, sizeof(u32),
796 					 &ap->evt_prd_dma, GFP_KERNEL);
797 	if (ap->evt_prd == NULL)
798 		goto fail;
799 
800 	ap->rx_ret_prd = dma_alloc_coherent(&ap->pdev->dev, sizeof(u32),
801 					    &ap->rx_ret_prd_dma, GFP_KERNEL);
802 	if (ap->rx_ret_prd == NULL)
803 		goto fail;
804 
805 	ap->tx_csm = dma_alloc_coherent(&ap->pdev->dev, sizeof(u32),
806 					&ap->tx_csm_dma, GFP_KERNEL);
807 	if (ap->tx_csm == NULL)
808 		goto fail;
809 
810 	return 0;
811 
812 fail:
813 	/* Clean up. */
814 	ace_init_cleanup(dev);
815 	return 1;
816 }
817 
818 
819 /*
820  * Generic cleanup handling data allocated during init. Used when the
821  * module is unloaded or if an error occurs during initialization
822  */
823 static void ace_init_cleanup(struct net_device *dev)
824 {
825 	struct ace_private *ap;
826 
827 	ap = netdev_priv(dev);
828 
829 	ace_free_descriptors(dev);
830 
831 	if (ap->info)
832 		dma_free_coherent(&ap->pdev->dev, sizeof(struct ace_info),
833 				  ap->info, ap->info_dma);
834 	kfree(ap->skb);
835 	kfree(ap->trace_buf);
836 
837 	if (dev->irq)
838 		free_irq(dev->irq, dev);
839 
840 	iounmap(ap->regs);
841 }
842 
843 
844 /*
845  * Commands are considered to be slow.
846  */
847 static inline void ace_issue_cmd(struct ace_regs __iomem *regs, struct cmd *cmd)
848 {
849 	u32 idx;
850 
851 	idx = readl(&regs->CmdPrd);
852 
853 	writel(*(u32 *)(cmd), &regs->CmdRng[idx]);
854 	idx = (idx + 1) % CMD_RING_ENTRIES;
855 
856 	writel(idx, &regs->CmdPrd);
857 }
858 
859 
860 static int ace_init(struct net_device *dev)
861 {
862 	struct ace_private *ap;
863 	struct ace_regs __iomem *regs;
864 	struct ace_info *info = NULL;
865 	struct pci_dev *pdev;
866 	unsigned long myjif;
867 	u64 tmp_ptr;
868 	u32 tig_ver, mac1, mac2, tmp, pci_state;
869 	int board_idx, ecode = 0;
870 	short i;
871 	unsigned char cache_size;
872 	u8 addr[ETH_ALEN];
873 
874 	ap = netdev_priv(dev);
875 	regs = ap->regs;
876 
877 	board_idx = ap->board_idx;
878 
879 	/*
880 	 * aman@sgi.com - its useful to do a NIC reset here to
881 	 * address the `Firmware not running' problem subsequent
882 	 * to any crashes involving the NIC
883 	 */
884 	writel(HW_RESET | (HW_RESET << 24), &regs->HostCtrl);
885 	readl(&regs->HostCtrl);		/* PCI write posting */
886 	udelay(5);
887 
888 	/*
889 	 * Don't access any other registers before this point!
890 	 */
891 #ifdef __BIG_ENDIAN
892 	/*
893 	 * This will most likely need BYTE_SWAP once we switch
894 	 * to using __raw_writel()
895 	 */
896 	writel((WORD_SWAP | CLR_INT | ((WORD_SWAP | CLR_INT) << 24)),
897 	       &regs->HostCtrl);
898 #else
899 	writel((CLR_INT | WORD_SWAP | ((CLR_INT | WORD_SWAP) << 24)),
900 	       &regs->HostCtrl);
901 #endif
902 	readl(&regs->HostCtrl);		/* PCI write posting */
903 
904 	/*
905 	 * Stop the NIC CPU and clear pending interrupts
906 	 */
907 	writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
908 	readl(&regs->CpuCtrl);		/* PCI write posting */
909 	writel(0, &regs->Mb0Lo);
910 
911 	tig_ver = readl(&regs->HostCtrl) >> 28;
912 
913 	switch(tig_ver){
914 #ifndef CONFIG_ACENIC_OMIT_TIGON_I
915 	case 4:
916 	case 5:
917 		printk(KERN_INFO "  Tigon I  (Rev. %i), Firmware: %i.%i.%i, ",
918 		       tig_ver, ap->firmware_major, ap->firmware_minor,
919 		       ap->firmware_fix);
920 		writel(0, &regs->LocalCtrl);
921 		ap->version = 1;
922 		ap->tx_ring_entries = TIGON_I_TX_RING_ENTRIES;
923 		break;
924 #endif
925 	case 6:
926 		printk(KERN_INFO "  Tigon II (Rev. %i), Firmware: %i.%i.%i, ",
927 		       tig_ver, ap->firmware_major, ap->firmware_minor,
928 		       ap->firmware_fix);
929 		writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
930 		readl(&regs->CpuBCtrl);		/* PCI write posting */
931 		/*
932 		 * The SRAM bank size does _not_ indicate the amount
933 		 * of memory on the card, it controls the _bank_ size!
934 		 * Ie. a 1MB AceNIC will have two banks of 512KB.
935 		 */
936 		writel(SRAM_BANK_512K, &regs->LocalCtrl);
937 		writel(SYNC_SRAM_TIMING, &regs->MiscCfg);
938 		ap->version = 2;
939 		ap->tx_ring_entries = MAX_TX_RING_ENTRIES;
940 		break;
941 	default:
942 		printk(KERN_WARNING "  Unsupported Tigon version detected "
943 		       "(%i)\n", tig_ver);
944 		ecode = -ENODEV;
945 		goto init_error;
946 	}
947 
948 	/*
949 	 * ModeStat _must_ be set after the SRAM settings as this change
950 	 * seems to corrupt the ModeStat and possible other registers.
951 	 * The SRAM settings survive resets and setting it to the same
952 	 * value a second time works as well. This is what caused the
953 	 * `Firmware not running' problem on the Tigon II.
954 	 */
955 #ifdef __BIG_ENDIAN
956 	writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL | ACE_BYTE_SWAP_BD |
957 	       ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
958 #else
959 	writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL |
960 	       ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
961 #endif
962 	readl(&regs->ModeStat);		/* PCI write posting */
963 
964 	mac1 = 0;
965 	for(i = 0; i < 4; i++) {
966 		int t;
967 
968 		mac1 = mac1 << 8;
969 		t = read_eeprom_byte(dev, 0x8c+i);
970 		if (t < 0) {
971 			ecode = -EIO;
972 			goto init_error;
973 		} else
974 			mac1 |= (t & 0xff);
975 	}
976 	mac2 = 0;
977 	for(i = 4; i < 8; i++) {
978 		int t;
979 
980 		mac2 = mac2 << 8;
981 		t = read_eeprom_byte(dev, 0x8c+i);
982 		if (t < 0) {
983 			ecode = -EIO;
984 			goto init_error;
985 		} else
986 			mac2 |= (t & 0xff);
987 	}
988 
989 	writel(mac1, &regs->MacAddrHi);
990 	writel(mac2, &regs->MacAddrLo);
991 
992 	addr[0] = (mac1 >> 8) & 0xff;
993 	addr[1] = mac1 & 0xff;
994 	addr[2] = (mac2 >> 24) & 0xff;
995 	addr[3] = (mac2 >> 16) & 0xff;
996 	addr[4] = (mac2 >> 8) & 0xff;
997 	addr[5] = mac2 & 0xff;
998 	eth_hw_addr_set(dev, addr);
999 
1000 	printk("MAC: %pM\n", dev->dev_addr);
1001 
1002 	/*
1003 	 * Looks like this is necessary to deal with on all architectures,
1004 	 * even this %$#%$# N440BX Intel based thing doesn't get it right.
1005 	 * Ie. having two NICs in the machine, one will have the cache
1006 	 * line set at boot time, the other will not.
1007 	 */
1008 	pdev = ap->pdev;
1009 	pci_read_config_byte(pdev, PCI_CACHE_LINE_SIZE, &cache_size);
1010 	cache_size <<= 2;
1011 	if (cache_size != SMP_CACHE_BYTES) {
1012 		printk(KERN_INFO "  PCI cache line size set incorrectly "
1013 		       "(%i bytes) by BIOS/FW, ", cache_size);
1014 		if (cache_size > SMP_CACHE_BYTES)
1015 			printk("expecting %i\n", SMP_CACHE_BYTES);
1016 		else {
1017 			printk("correcting to %i\n", SMP_CACHE_BYTES);
1018 			pci_write_config_byte(pdev, PCI_CACHE_LINE_SIZE,
1019 					      SMP_CACHE_BYTES >> 2);
1020 		}
1021 	}
1022 
1023 	pci_state = readl(&regs->PciState);
1024 	printk(KERN_INFO "  PCI bus width: %i bits, speed: %iMHz, "
1025 	       "latency: %i clks\n",
1026 	       	(pci_state & PCI_32BIT) ? 32 : 64,
1027 		(pci_state & PCI_66MHZ) ? 66 : 33,
1028 		ap->pci_latency);
1029 
1030 	/*
1031 	 * Set the max DMA transfer size. Seems that for most systems
1032 	 * the performance is better when no MAX parameter is
1033 	 * set. However for systems enabling PCI write and invalidate,
1034 	 * DMA writes must be set to the L1 cache line size to get
1035 	 * optimal performance.
1036 	 *
1037 	 * The default is now to turn the PCI write and invalidate off
1038 	 * - that is what Alteon does for NT.
1039 	 */
1040 	tmp = READ_CMD_MEM | WRITE_CMD_MEM;
1041 	if (ap->version >= 2) {
1042 		tmp |= (MEM_READ_MULTIPLE | (pci_state & PCI_66MHZ));
1043 		/*
1044 		 * Tuning parameters only supported for 8 cards
1045 		 */
1046 		if (board_idx == BOARD_IDX_OVERFLOW ||
1047 		    dis_pci_mem_inval[board_idx]) {
1048 			if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1049 				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1050 				pci_write_config_word(pdev, PCI_COMMAND,
1051 						      ap->pci_command);
1052 				printk(KERN_INFO "  Disabling PCI memory "
1053 				       "write and invalidate\n");
1054 			}
1055 		} else if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1056 			printk(KERN_INFO "  PCI memory write & invalidate "
1057 			       "enabled by BIOS, enabling counter measures\n");
1058 
1059 			switch(SMP_CACHE_BYTES) {
1060 			case 16:
1061 				tmp |= DMA_WRITE_MAX_16;
1062 				break;
1063 			case 32:
1064 				tmp |= DMA_WRITE_MAX_32;
1065 				break;
1066 			case 64:
1067 				tmp |= DMA_WRITE_MAX_64;
1068 				break;
1069 			case 128:
1070 				tmp |= DMA_WRITE_MAX_128;
1071 				break;
1072 			default:
1073 				printk(KERN_INFO "  Cache line size %i not "
1074 				       "supported, PCI write and invalidate "
1075 				       "disabled\n", SMP_CACHE_BYTES);
1076 				ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1077 				pci_write_config_word(pdev, PCI_COMMAND,
1078 						      ap->pci_command);
1079 			}
1080 		}
1081 	}
1082 
1083 #ifdef __sparc__
1084 	/*
1085 	 * On this platform, we know what the best dma settings
1086 	 * are.  We use 64-byte maximum bursts, because if we
1087 	 * burst larger than the cache line size (or even cross
1088 	 * a 64byte boundary in a single burst) the UltraSparc
1089 	 * PCI controller will disconnect at 64-byte multiples.
1090 	 *
1091 	 * Read-multiple will be properly enabled above, and when
1092 	 * set will give the PCI controller proper hints about
1093 	 * prefetching.
1094 	 */
1095 	tmp &= ~DMA_READ_WRITE_MASK;
1096 	tmp |= DMA_READ_MAX_64;
1097 	tmp |= DMA_WRITE_MAX_64;
1098 #endif
1099 #ifdef __alpha__
1100 	tmp &= ~DMA_READ_WRITE_MASK;
1101 	tmp |= DMA_READ_MAX_128;
1102 	/*
1103 	 * All the docs say MUST NOT. Well, I did.
1104 	 * Nothing terrible happens, if we load wrong size.
1105 	 * Bit w&i still works better!
1106 	 */
1107 	tmp |= DMA_WRITE_MAX_128;
1108 #endif
1109 	writel(tmp, &regs->PciState);
1110 
1111 #if 0
1112 	/*
1113 	 * The Host PCI bus controller driver has to set FBB.
1114 	 * If all devices on that PCI bus support FBB, then the controller
1115 	 * can enable FBB support in the Host PCI Bus controller (or on
1116 	 * the PCI-PCI bridge if that applies).
1117 	 * -ggg
1118 	 */
1119 	/*
1120 	 * I have received reports from people having problems when this
1121 	 * bit is enabled.
1122 	 */
1123 	if (!(ap->pci_command & PCI_COMMAND_FAST_BACK)) {
1124 		printk(KERN_INFO "  Enabling PCI Fast Back to Back\n");
1125 		ap->pci_command |= PCI_COMMAND_FAST_BACK;
1126 		pci_write_config_word(pdev, PCI_COMMAND, ap->pci_command);
1127 	}
1128 #endif
1129 
1130 	/*
1131 	 * Configure DMA attributes.
1132 	 */
1133 	if (!dma_set_mask(&pdev->dev, DMA_BIT_MASK(64))) {
1134 		ap->pci_using_dac = 1;
1135 	} else if (!dma_set_mask(&pdev->dev, DMA_BIT_MASK(32))) {
1136 		ap->pci_using_dac = 0;
1137 	} else {
1138 		ecode = -ENODEV;
1139 		goto init_error;
1140 	}
1141 
1142 	/*
1143 	 * Initialize the generic info block and the command+event rings
1144 	 * and the control blocks for the transmit and receive rings
1145 	 * as they need to be setup once and for all.
1146 	 */
1147 	if (!(info = dma_alloc_coherent(&ap->pdev->dev, sizeof(struct ace_info),
1148 					&ap->info_dma, GFP_KERNEL))) {
1149 		ecode = -EAGAIN;
1150 		goto init_error;
1151 	}
1152 	ap->info = info;
1153 
1154 	/*
1155 	 * Get the memory for the skb rings.
1156 	 */
1157 	if (!(ap->skb = kzalloc(sizeof(struct ace_skb), GFP_KERNEL))) {
1158 		ecode = -EAGAIN;
1159 		goto init_error;
1160 	}
1161 
1162 	ecode = request_irq(pdev->irq, ace_interrupt, IRQF_SHARED,
1163 			    DRV_NAME, dev);
1164 	if (ecode) {
1165 		printk(KERN_WARNING "%s: Requested IRQ %d is busy\n",
1166 		       DRV_NAME, pdev->irq);
1167 		goto init_error;
1168 	} else
1169 		dev->irq = pdev->irq;
1170 
1171 #ifdef INDEX_DEBUG
1172 	spin_lock_init(&ap->debug_lock);
1173 	ap->last_tx = ACE_TX_RING_ENTRIES(ap) - 1;
1174 	ap->last_std_rx = 0;
1175 	ap->last_mini_rx = 0;
1176 #endif
1177 
1178 	ecode = ace_load_firmware(dev);
1179 	if (ecode)
1180 		goto init_error;
1181 
1182 	ap->fw_running = 0;
1183 
1184 	tmp_ptr = ap->info_dma;
1185 	writel(tmp_ptr >> 32, &regs->InfoPtrHi);
1186 	writel(tmp_ptr & 0xffffffff, &regs->InfoPtrLo);
1187 
1188 	memset(ap->evt_ring, 0, EVT_RING_ENTRIES * sizeof(struct event));
1189 
1190 	set_aceaddr(&info->evt_ctrl.rngptr, ap->evt_ring_dma);
1191 	info->evt_ctrl.flags = 0;
1192 
1193 	*(ap->evt_prd) = 0;
1194 	wmb();
1195 	set_aceaddr(&info->evt_prd_ptr, ap->evt_prd_dma);
1196 	writel(0, &regs->EvtCsm);
1197 
1198 	set_aceaddr(&info->cmd_ctrl.rngptr, 0x100);
1199 	info->cmd_ctrl.flags = 0;
1200 	info->cmd_ctrl.max_len = 0;
1201 
1202 	for (i = 0; i < CMD_RING_ENTRIES; i++)
1203 		writel(0, &regs->CmdRng[i]);
1204 
1205 	writel(0, &regs->CmdPrd);
1206 	writel(0, &regs->CmdCsm);
1207 
1208 	tmp_ptr = ap->info_dma;
1209 	tmp_ptr += (unsigned long) &(((struct ace_info *)0)->s.stats);
1210 	set_aceaddr(&info->stats2_ptr, (dma_addr_t) tmp_ptr);
1211 
1212 	set_aceaddr(&info->rx_std_ctrl.rngptr, ap->rx_ring_base_dma);
1213 	info->rx_std_ctrl.max_len = ACE_STD_BUFSIZE;
1214 	info->rx_std_ctrl.flags =
1215 	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1216 
1217 	memset(ap->rx_std_ring, 0,
1218 	       RX_STD_RING_ENTRIES * sizeof(struct rx_desc));
1219 
1220 	for (i = 0; i < RX_STD_RING_ENTRIES; i++)
1221 		ap->rx_std_ring[i].flags = BD_FLG_TCP_UDP_SUM;
1222 
1223 	ap->rx_std_skbprd = 0;
1224 	atomic_set(&ap->cur_rx_bufs, 0);
1225 
1226 	set_aceaddr(&info->rx_jumbo_ctrl.rngptr,
1227 		    (ap->rx_ring_base_dma +
1228 		     (sizeof(struct rx_desc) * RX_STD_RING_ENTRIES)));
1229 	info->rx_jumbo_ctrl.max_len = 0;
1230 	info->rx_jumbo_ctrl.flags =
1231 	  RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1232 
1233 	memset(ap->rx_jumbo_ring, 0,
1234 	       RX_JUMBO_RING_ENTRIES * sizeof(struct rx_desc));
1235 
1236 	for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++)
1237 		ap->rx_jumbo_ring[i].flags = BD_FLG_TCP_UDP_SUM | BD_FLG_JUMBO;
1238 
1239 	ap->rx_jumbo_skbprd = 0;
1240 	atomic_set(&ap->cur_jumbo_bufs, 0);
1241 
1242 	memset(ap->rx_mini_ring, 0,
1243 	       RX_MINI_RING_ENTRIES * sizeof(struct rx_desc));
1244 
1245 	if (ap->version >= 2) {
1246 		set_aceaddr(&info->rx_mini_ctrl.rngptr,
1247 			    (ap->rx_ring_base_dma +
1248 			     (sizeof(struct rx_desc) *
1249 			      (RX_STD_RING_ENTRIES +
1250 			       RX_JUMBO_RING_ENTRIES))));
1251 		info->rx_mini_ctrl.max_len = ACE_MINI_SIZE;
1252 		info->rx_mini_ctrl.flags =
1253 		  RCB_FLG_TCP_UDP_SUM|RCB_FLG_NO_PSEUDO_HDR|RCB_FLG_VLAN_ASSIST;
1254 
1255 		for (i = 0; i < RX_MINI_RING_ENTRIES; i++)
1256 			ap->rx_mini_ring[i].flags =
1257 				BD_FLG_TCP_UDP_SUM | BD_FLG_MINI;
1258 	} else {
1259 		set_aceaddr(&info->rx_mini_ctrl.rngptr, 0);
1260 		info->rx_mini_ctrl.flags = RCB_FLG_RNG_DISABLE;
1261 		info->rx_mini_ctrl.max_len = 0;
1262 	}
1263 
1264 	ap->rx_mini_skbprd = 0;
1265 	atomic_set(&ap->cur_mini_bufs, 0);
1266 
1267 	set_aceaddr(&info->rx_return_ctrl.rngptr,
1268 		    (ap->rx_ring_base_dma +
1269 		     (sizeof(struct rx_desc) *
1270 		      (RX_STD_RING_ENTRIES +
1271 		       RX_JUMBO_RING_ENTRIES +
1272 		       RX_MINI_RING_ENTRIES))));
1273 	info->rx_return_ctrl.flags = 0;
1274 	info->rx_return_ctrl.max_len = RX_RETURN_RING_ENTRIES;
1275 
1276 	memset(ap->rx_return_ring, 0,
1277 	       RX_RETURN_RING_ENTRIES * sizeof(struct rx_desc));
1278 
1279 	set_aceaddr(&info->rx_ret_prd_ptr, ap->rx_ret_prd_dma);
1280 	*(ap->rx_ret_prd) = 0;
1281 
1282 	writel(TX_RING_BASE, &regs->WinBase);
1283 
1284 	if (ACE_IS_TIGON_I(ap)) {
1285 		ap->tx_ring = (__force struct tx_desc *) regs->Window;
1286 		for (i = 0; i < (TIGON_I_TX_RING_ENTRIES
1287 				 * sizeof(struct tx_desc)) / sizeof(u32); i++)
1288 			writel(0, (__force void __iomem *)ap->tx_ring  + i * 4);
1289 
1290 		set_aceaddr(&info->tx_ctrl.rngptr, TX_RING_BASE);
1291 	} else {
1292 		memset(ap->tx_ring, 0,
1293 		       MAX_TX_RING_ENTRIES * sizeof(struct tx_desc));
1294 
1295 		set_aceaddr(&info->tx_ctrl.rngptr, ap->tx_ring_dma);
1296 	}
1297 
1298 	info->tx_ctrl.max_len = ACE_TX_RING_ENTRIES(ap);
1299 	tmp = RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | RCB_FLG_VLAN_ASSIST;
1300 
1301 	/*
1302 	 * The Tigon I does not like having the TX ring in host memory ;-(
1303 	 */
1304 	if (!ACE_IS_TIGON_I(ap))
1305 		tmp |= RCB_FLG_TX_HOST_RING;
1306 #if TX_COAL_INTS_ONLY
1307 	tmp |= RCB_FLG_COAL_INT_ONLY;
1308 #endif
1309 	info->tx_ctrl.flags = tmp;
1310 
1311 	set_aceaddr(&info->tx_csm_ptr, ap->tx_csm_dma);
1312 
1313 	/*
1314 	 * Potential item for tuning parameter
1315 	 */
1316 #if 0 /* NO */
1317 	writel(DMA_THRESH_16W, &regs->DmaReadCfg);
1318 	writel(DMA_THRESH_16W, &regs->DmaWriteCfg);
1319 #else
1320 	writel(DMA_THRESH_8W, &regs->DmaReadCfg);
1321 	writel(DMA_THRESH_8W, &regs->DmaWriteCfg);
1322 #endif
1323 
1324 	writel(0, &regs->MaskInt);
1325 	writel(1, &regs->IfIdx);
1326 #if 0
1327 	/*
1328 	 * McKinley boxes do not like us fiddling with AssistState
1329 	 * this early
1330 	 */
1331 	writel(1, &regs->AssistState);
1332 #endif
1333 
1334 	writel(DEF_STAT, &regs->TuneStatTicks);
1335 	writel(DEF_TRACE, &regs->TuneTrace);
1336 
1337 	ace_set_rxtx_parms(dev, 0);
1338 
1339 	if (board_idx == BOARD_IDX_OVERFLOW) {
1340 		printk(KERN_WARNING "%s: more than %i NICs detected, "
1341 		       "ignoring module parameters!\n",
1342 		       ap->name, ACE_MAX_MOD_PARMS);
1343 	} else if (board_idx >= 0) {
1344 		if (tx_coal_tick[board_idx])
1345 			writel(tx_coal_tick[board_idx],
1346 			       &regs->TuneTxCoalTicks);
1347 		if (max_tx_desc[board_idx])
1348 			writel(max_tx_desc[board_idx], &regs->TuneMaxTxDesc);
1349 
1350 		if (rx_coal_tick[board_idx])
1351 			writel(rx_coal_tick[board_idx],
1352 			       &regs->TuneRxCoalTicks);
1353 		if (max_rx_desc[board_idx])
1354 			writel(max_rx_desc[board_idx], &regs->TuneMaxRxDesc);
1355 
1356 		if (trace[board_idx])
1357 			writel(trace[board_idx], &regs->TuneTrace);
1358 
1359 		if ((tx_ratio[board_idx] > 0) && (tx_ratio[board_idx] < 64))
1360 			writel(tx_ratio[board_idx], &regs->TxBufRat);
1361 	}
1362 
1363 	/*
1364 	 * Default link parameters
1365 	 */
1366 	tmp = LNK_ENABLE | LNK_FULL_DUPLEX | LNK_1000MB | LNK_100MB |
1367 		LNK_10MB | LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL | LNK_NEGOTIATE;
1368 	if(ap->version >= 2)
1369 		tmp |= LNK_TX_FLOW_CTL_Y;
1370 
1371 	/*
1372 	 * Override link default parameters
1373 	 */
1374 	if ((board_idx >= 0) && link_state[board_idx]) {
1375 		int option = link_state[board_idx];
1376 
1377 		tmp = LNK_ENABLE;
1378 
1379 		if (option & 0x01) {
1380 			printk(KERN_INFO "%s: Setting half duplex link\n",
1381 			       ap->name);
1382 			tmp &= ~LNK_FULL_DUPLEX;
1383 		}
1384 		if (option & 0x02)
1385 			tmp &= ~LNK_NEGOTIATE;
1386 		if (option & 0x10)
1387 			tmp |= LNK_10MB;
1388 		if (option & 0x20)
1389 			tmp |= LNK_100MB;
1390 		if (option & 0x40)
1391 			tmp |= LNK_1000MB;
1392 		if ((option & 0x70) == 0) {
1393 			printk(KERN_WARNING "%s: No media speed specified, "
1394 			       "forcing auto negotiation\n", ap->name);
1395 			tmp |= LNK_NEGOTIATE | LNK_1000MB |
1396 				LNK_100MB | LNK_10MB;
1397 		}
1398 		if ((option & 0x100) == 0)
1399 			tmp |= LNK_NEG_FCTL;
1400 		else
1401 			printk(KERN_INFO "%s: Disabling flow control "
1402 			       "negotiation\n", ap->name);
1403 		if (option & 0x200)
1404 			tmp |= LNK_RX_FLOW_CTL_Y;
1405 		if ((option & 0x400) && (ap->version >= 2)) {
1406 			printk(KERN_INFO "%s: Enabling TX flow control\n",
1407 			       ap->name);
1408 			tmp |= LNK_TX_FLOW_CTL_Y;
1409 		}
1410 	}
1411 
1412 	ap->link = tmp;
1413 	writel(tmp, &regs->TuneLink);
1414 	if (ap->version >= 2)
1415 		writel(tmp, &regs->TuneFastLink);
1416 
1417 	writel(ap->firmware_start, &regs->Pc);
1418 
1419 	writel(0, &regs->Mb0Lo);
1420 
1421 	/*
1422 	 * Set tx_csm before we start receiving interrupts, otherwise
1423 	 * the interrupt handler might think it is supposed to process
1424 	 * tx ints before we are up and running, which may cause a null
1425 	 * pointer access in the int handler.
1426 	 */
1427 	ap->cur_rx = 0;
1428 	ap->tx_prd = *(ap->tx_csm) = ap->tx_ret_csm = 0;
1429 
1430 	wmb();
1431 	ace_set_txprd(regs, ap, 0);
1432 	writel(0, &regs->RxRetCsm);
1433 
1434 	/*
1435 	 * Enable DMA engine now.
1436 	 * If we do this sooner, Mckinley box pukes.
1437 	 * I assume it's because Tigon II DMA engine wants to check
1438 	 * *something* even before the CPU is started.
1439 	 */
1440 	writel(1, &regs->AssistState);  /* enable DMA */
1441 
1442 	/*
1443 	 * Start the NIC CPU
1444 	 */
1445 	writel(readl(&regs->CpuCtrl) & ~(CPU_HALT|CPU_TRACE), &regs->CpuCtrl);
1446 	readl(&regs->CpuCtrl);
1447 
1448 	/*
1449 	 * Wait for the firmware to spin up - max 3 seconds.
1450 	 */
1451 	myjif = jiffies + 3 * HZ;
1452 	while (time_before(jiffies, myjif) && !ap->fw_running)
1453 		cpu_relax();
1454 
1455 	if (!ap->fw_running) {
1456 		printk(KERN_ERR "%s: Firmware NOT running!\n", ap->name);
1457 
1458 		ace_dump_trace(ap);
1459 		writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
1460 		readl(&regs->CpuCtrl);
1461 
1462 		/* aman@sgi.com - account for badly behaving firmware/NIC:
1463 		 * - have observed that the NIC may continue to generate
1464 		 *   interrupts for some reason; attempt to stop it - halt
1465 		 *   second CPU for Tigon II cards, and also clear Mb0
1466 		 * - if we're a module, we'll fail to load if this was
1467 		 *   the only GbE card in the system => if the kernel does
1468 		 *   see an interrupt from the NIC, code to handle it is
1469 		 *   gone and OOps! - so free_irq also
1470 		 */
1471 		if (ap->version >= 2)
1472 			writel(readl(&regs->CpuBCtrl) | CPU_HALT,
1473 			       &regs->CpuBCtrl);
1474 		writel(0, &regs->Mb0Lo);
1475 		readl(&regs->Mb0Lo);
1476 
1477 		ecode = -EBUSY;
1478 		goto init_error;
1479 	}
1480 
1481 	/*
1482 	 * We load the ring here as there seem to be no way to tell the
1483 	 * firmware to wipe the ring without re-initializing it.
1484 	 */
1485 	if (!test_and_set_bit(0, &ap->std_refill_busy))
1486 		ace_load_std_rx_ring(dev, RX_RING_SIZE);
1487 	else
1488 		printk(KERN_ERR "%s: Someone is busy refilling the RX ring\n",
1489 		       ap->name);
1490 	if (ap->version >= 2) {
1491 		if (!test_and_set_bit(0, &ap->mini_refill_busy))
1492 			ace_load_mini_rx_ring(dev, RX_MINI_SIZE);
1493 		else
1494 			printk(KERN_ERR "%s: Someone is busy refilling "
1495 			       "the RX mini ring\n", ap->name);
1496 	}
1497 	return 0;
1498 
1499  init_error:
1500 	ace_init_cleanup(dev);
1501 	return ecode;
1502 }
1503 
1504 
1505 static void ace_set_rxtx_parms(struct net_device *dev, int jumbo)
1506 {
1507 	struct ace_private *ap = netdev_priv(dev);
1508 	struct ace_regs __iomem *regs = ap->regs;
1509 	int board_idx = ap->board_idx;
1510 
1511 	if (board_idx >= 0) {
1512 		if (!jumbo) {
1513 			if (!tx_coal_tick[board_idx])
1514 				writel(DEF_TX_COAL, &regs->TuneTxCoalTicks);
1515 			if (!max_tx_desc[board_idx])
1516 				writel(DEF_TX_MAX_DESC, &regs->TuneMaxTxDesc);
1517 			if (!rx_coal_tick[board_idx])
1518 				writel(DEF_RX_COAL, &regs->TuneRxCoalTicks);
1519 			if (!max_rx_desc[board_idx])
1520 				writel(DEF_RX_MAX_DESC, &regs->TuneMaxRxDesc);
1521 			if (!tx_ratio[board_idx])
1522 				writel(DEF_TX_RATIO, &regs->TxBufRat);
1523 		} else {
1524 			if (!tx_coal_tick[board_idx])
1525 				writel(DEF_JUMBO_TX_COAL,
1526 				       &regs->TuneTxCoalTicks);
1527 			if (!max_tx_desc[board_idx])
1528 				writel(DEF_JUMBO_TX_MAX_DESC,
1529 				       &regs->TuneMaxTxDesc);
1530 			if (!rx_coal_tick[board_idx])
1531 				writel(DEF_JUMBO_RX_COAL,
1532 				       &regs->TuneRxCoalTicks);
1533 			if (!max_rx_desc[board_idx])
1534 				writel(DEF_JUMBO_RX_MAX_DESC,
1535 				       &regs->TuneMaxRxDesc);
1536 			if (!tx_ratio[board_idx])
1537 				writel(DEF_JUMBO_TX_RATIO, &regs->TxBufRat);
1538 		}
1539 	}
1540 }
1541 
1542 
1543 static void ace_watchdog(struct net_device *data, unsigned int txqueue)
1544 {
1545 	struct net_device *dev = data;
1546 	struct ace_private *ap = netdev_priv(dev);
1547 	struct ace_regs __iomem *regs = ap->regs;
1548 
1549 	/*
1550 	 * We haven't received a stats update event for more than 2.5
1551 	 * seconds and there is data in the transmit queue, thus we
1552 	 * assume the card is stuck.
1553 	 */
1554 	if (*ap->tx_csm != ap->tx_ret_csm) {
1555 		printk(KERN_WARNING "%s: Transmitter is stuck, %08x\n",
1556 		       dev->name, (unsigned int)readl(&regs->HostCtrl));
1557 		/* This can happen due to ieee flow control. */
1558 	} else {
1559 		printk(KERN_DEBUG "%s: BUG... transmitter died. Kicking it.\n",
1560 		       dev->name);
1561 #if 0
1562 		netif_wake_queue(dev);
1563 #endif
1564 	}
1565 }
1566 
1567 
1568 static void ace_tasklet(struct tasklet_struct *t)
1569 {
1570 	struct ace_private *ap = from_tasklet(ap, t, ace_tasklet);
1571 	struct net_device *dev = ap->ndev;
1572 	int cur_size;
1573 
1574 	cur_size = atomic_read(&ap->cur_rx_bufs);
1575 	if ((cur_size < RX_LOW_STD_THRES) &&
1576 	    !test_and_set_bit(0, &ap->std_refill_busy)) {
1577 #ifdef DEBUG
1578 		printk("refilling buffers (current %i)\n", cur_size);
1579 #endif
1580 		ace_load_std_rx_ring(dev, RX_RING_SIZE - cur_size);
1581 	}
1582 
1583 	if (ap->version >= 2) {
1584 		cur_size = atomic_read(&ap->cur_mini_bufs);
1585 		if ((cur_size < RX_LOW_MINI_THRES) &&
1586 		    !test_and_set_bit(0, &ap->mini_refill_busy)) {
1587 #ifdef DEBUG
1588 			printk("refilling mini buffers (current %i)\n",
1589 			       cur_size);
1590 #endif
1591 			ace_load_mini_rx_ring(dev, RX_MINI_SIZE - cur_size);
1592 		}
1593 	}
1594 
1595 	cur_size = atomic_read(&ap->cur_jumbo_bufs);
1596 	if (ap->jumbo && (cur_size < RX_LOW_JUMBO_THRES) &&
1597 	    !test_and_set_bit(0, &ap->jumbo_refill_busy)) {
1598 #ifdef DEBUG
1599 		printk("refilling jumbo buffers (current %i)\n", cur_size);
1600 #endif
1601 		ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE - cur_size);
1602 	}
1603 	ap->tasklet_pending = 0;
1604 }
1605 
1606 
1607 /*
1608  * Copy the contents of the NIC's trace buffer to kernel memory.
1609  */
1610 static void ace_dump_trace(struct ace_private *ap)
1611 {
1612 #if 0
1613 	if (!ap->trace_buf)
1614 		if (!(ap->trace_buf = kmalloc(ACE_TRACE_SIZE, GFP_KERNEL)))
1615 		    return;
1616 #endif
1617 }
1618 
1619 
1620 /*
1621  * Load the standard rx ring.
1622  *
1623  * Loading rings is safe without holding the spin lock since this is
1624  * done only before the device is enabled, thus no interrupts are
1625  * generated and by the interrupt handler/tasklet handler.
1626  */
1627 static void ace_load_std_rx_ring(struct net_device *dev, int nr_bufs)
1628 {
1629 	struct ace_private *ap = netdev_priv(dev);
1630 	struct ace_regs __iomem *regs = ap->regs;
1631 	short i, idx;
1632 
1633 
1634 	prefetchw(&ap->cur_rx_bufs);
1635 
1636 	idx = ap->rx_std_skbprd;
1637 
1638 	for (i = 0; i < nr_bufs; i++) {
1639 		struct sk_buff *skb;
1640 		struct rx_desc *rd;
1641 		dma_addr_t mapping;
1642 
1643 		skb = netdev_alloc_skb_ip_align(dev, ACE_STD_BUFSIZE);
1644 		if (!skb)
1645 			break;
1646 
1647 		mapping = dma_map_page(&ap->pdev->dev,
1648 				       virt_to_page(skb->data),
1649 				       offset_in_page(skb->data),
1650 				       ACE_STD_BUFSIZE, DMA_FROM_DEVICE);
1651 		ap->skb->rx_std_skbuff[idx].skb = skb;
1652 		dma_unmap_addr_set(&ap->skb->rx_std_skbuff[idx],
1653 				   mapping, mapping);
1654 
1655 		rd = &ap->rx_std_ring[idx];
1656 		set_aceaddr(&rd->addr, mapping);
1657 		rd->size = ACE_STD_BUFSIZE;
1658 		rd->idx = idx;
1659 		idx = (idx + 1) % RX_STD_RING_ENTRIES;
1660 	}
1661 
1662 	if (!i)
1663 		goto error_out;
1664 
1665 	atomic_add(i, &ap->cur_rx_bufs);
1666 	ap->rx_std_skbprd = idx;
1667 
1668 	if (ACE_IS_TIGON_I(ap)) {
1669 		struct cmd cmd;
1670 		cmd.evt = C_SET_RX_PRD_IDX;
1671 		cmd.code = 0;
1672 		cmd.idx = ap->rx_std_skbprd;
1673 		ace_issue_cmd(regs, &cmd);
1674 	} else {
1675 		writel(idx, &regs->RxStdPrd);
1676 		wmb();
1677 	}
1678 
1679  out:
1680 	clear_bit(0, &ap->std_refill_busy);
1681 	return;
1682 
1683  error_out:
1684 	printk(KERN_INFO "Out of memory when allocating "
1685 	       "standard receive buffers\n");
1686 	goto out;
1687 }
1688 
1689 
1690 static void ace_load_mini_rx_ring(struct net_device *dev, int nr_bufs)
1691 {
1692 	struct ace_private *ap = netdev_priv(dev);
1693 	struct ace_regs __iomem *regs = ap->regs;
1694 	short i, idx;
1695 
1696 	prefetchw(&ap->cur_mini_bufs);
1697 
1698 	idx = ap->rx_mini_skbprd;
1699 	for (i = 0; i < nr_bufs; i++) {
1700 		struct sk_buff *skb;
1701 		struct rx_desc *rd;
1702 		dma_addr_t mapping;
1703 
1704 		skb = netdev_alloc_skb_ip_align(dev, ACE_MINI_BUFSIZE);
1705 		if (!skb)
1706 			break;
1707 
1708 		mapping = dma_map_page(&ap->pdev->dev,
1709 				       virt_to_page(skb->data),
1710 				       offset_in_page(skb->data),
1711 				       ACE_MINI_BUFSIZE, DMA_FROM_DEVICE);
1712 		ap->skb->rx_mini_skbuff[idx].skb = skb;
1713 		dma_unmap_addr_set(&ap->skb->rx_mini_skbuff[idx],
1714 				   mapping, mapping);
1715 
1716 		rd = &ap->rx_mini_ring[idx];
1717 		set_aceaddr(&rd->addr, mapping);
1718 		rd->size = ACE_MINI_BUFSIZE;
1719 		rd->idx = idx;
1720 		idx = (idx + 1) % RX_MINI_RING_ENTRIES;
1721 	}
1722 
1723 	if (!i)
1724 		goto error_out;
1725 
1726 	atomic_add(i, &ap->cur_mini_bufs);
1727 
1728 	ap->rx_mini_skbprd = idx;
1729 
1730 	writel(idx, &regs->RxMiniPrd);
1731 	wmb();
1732 
1733  out:
1734 	clear_bit(0, &ap->mini_refill_busy);
1735 	return;
1736  error_out:
1737 	printk(KERN_INFO "Out of memory when allocating "
1738 	       "mini receive buffers\n");
1739 	goto out;
1740 }
1741 
1742 
1743 /*
1744  * Load the jumbo rx ring, this may happen at any time if the MTU
1745  * is changed to a value > 1500.
1746  */
1747 static void ace_load_jumbo_rx_ring(struct net_device *dev, int nr_bufs)
1748 {
1749 	struct ace_private *ap = netdev_priv(dev);
1750 	struct ace_regs __iomem *regs = ap->regs;
1751 	short i, idx;
1752 
1753 	idx = ap->rx_jumbo_skbprd;
1754 
1755 	for (i = 0; i < nr_bufs; i++) {
1756 		struct sk_buff *skb;
1757 		struct rx_desc *rd;
1758 		dma_addr_t mapping;
1759 
1760 		skb = netdev_alloc_skb_ip_align(dev, ACE_JUMBO_BUFSIZE);
1761 		if (!skb)
1762 			break;
1763 
1764 		mapping = dma_map_page(&ap->pdev->dev,
1765 				       virt_to_page(skb->data),
1766 				       offset_in_page(skb->data),
1767 				       ACE_JUMBO_BUFSIZE, DMA_FROM_DEVICE);
1768 		ap->skb->rx_jumbo_skbuff[idx].skb = skb;
1769 		dma_unmap_addr_set(&ap->skb->rx_jumbo_skbuff[idx],
1770 				   mapping, mapping);
1771 
1772 		rd = &ap->rx_jumbo_ring[idx];
1773 		set_aceaddr(&rd->addr, mapping);
1774 		rd->size = ACE_JUMBO_BUFSIZE;
1775 		rd->idx = idx;
1776 		idx = (idx + 1) % RX_JUMBO_RING_ENTRIES;
1777 	}
1778 
1779 	if (!i)
1780 		goto error_out;
1781 
1782 	atomic_add(i, &ap->cur_jumbo_bufs);
1783 	ap->rx_jumbo_skbprd = idx;
1784 
1785 	if (ACE_IS_TIGON_I(ap)) {
1786 		struct cmd cmd;
1787 		cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1788 		cmd.code = 0;
1789 		cmd.idx = ap->rx_jumbo_skbprd;
1790 		ace_issue_cmd(regs, &cmd);
1791 	} else {
1792 		writel(idx, &regs->RxJumboPrd);
1793 		wmb();
1794 	}
1795 
1796  out:
1797 	clear_bit(0, &ap->jumbo_refill_busy);
1798 	return;
1799  error_out:
1800 	if (net_ratelimit())
1801 		printk(KERN_INFO "Out of memory when allocating "
1802 		       "jumbo receive buffers\n");
1803 	goto out;
1804 }
1805 
1806 
1807 /*
1808  * All events are considered to be slow (RX/TX ints do not generate
1809  * events) and are handled here, outside the main interrupt handler,
1810  * to reduce the size of the handler.
1811  */
1812 static u32 ace_handle_event(struct net_device *dev, u32 evtcsm, u32 evtprd)
1813 {
1814 	struct ace_private *ap;
1815 
1816 	ap = netdev_priv(dev);
1817 
1818 	while (evtcsm != evtprd) {
1819 		switch (ap->evt_ring[evtcsm].evt) {
1820 		case E_FW_RUNNING:
1821 			printk(KERN_INFO "%s: Firmware up and running\n",
1822 			       ap->name);
1823 			ap->fw_running = 1;
1824 			wmb();
1825 			break;
1826 		case E_STATS_UPDATED:
1827 			break;
1828 		case E_LNK_STATE:
1829 		{
1830 			u16 code = ap->evt_ring[evtcsm].code;
1831 			switch (code) {
1832 			case E_C_LINK_UP:
1833 			{
1834 				u32 state = readl(&ap->regs->GigLnkState);
1835 				printk(KERN_WARNING "%s: Optical link UP "
1836 				       "(%s Duplex, Flow Control: %s%s)\n",
1837 				       ap->name,
1838 				       state & LNK_FULL_DUPLEX ? "Full":"Half",
1839 				       state & LNK_TX_FLOW_CTL_Y ? "TX " : "",
1840 				       state & LNK_RX_FLOW_CTL_Y ? "RX" : "");
1841 				break;
1842 			}
1843 			case E_C_LINK_DOWN:
1844 				printk(KERN_WARNING "%s: Optical link DOWN\n",
1845 				       ap->name);
1846 				break;
1847 			case E_C_LINK_10_100:
1848 				printk(KERN_WARNING "%s: 10/100BaseT link "
1849 				       "UP\n", ap->name);
1850 				break;
1851 			default:
1852 				printk(KERN_ERR "%s: Unknown optical link "
1853 				       "state %02x\n", ap->name, code);
1854 			}
1855 			break;
1856 		}
1857 		case E_ERROR:
1858 			switch(ap->evt_ring[evtcsm].code) {
1859 			case E_C_ERR_INVAL_CMD:
1860 				printk(KERN_ERR "%s: invalid command error\n",
1861 				       ap->name);
1862 				break;
1863 			case E_C_ERR_UNIMP_CMD:
1864 				printk(KERN_ERR "%s: unimplemented command "
1865 				       "error\n", ap->name);
1866 				break;
1867 			case E_C_ERR_BAD_CFG:
1868 				printk(KERN_ERR "%s: bad config error\n",
1869 				       ap->name);
1870 				break;
1871 			default:
1872 				printk(KERN_ERR "%s: unknown error %02x\n",
1873 				       ap->name, ap->evt_ring[evtcsm].code);
1874 			}
1875 			break;
1876 		case E_RESET_JUMBO_RNG:
1877 		{
1878 			int i;
1879 			for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
1880 				if (ap->skb->rx_jumbo_skbuff[i].skb) {
1881 					ap->rx_jumbo_ring[i].size = 0;
1882 					set_aceaddr(&ap->rx_jumbo_ring[i].addr, 0);
1883 					dev_kfree_skb(ap->skb->rx_jumbo_skbuff[i].skb);
1884 					ap->skb->rx_jumbo_skbuff[i].skb = NULL;
1885 				}
1886 			}
1887 
1888 			if (ACE_IS_TIGON_I(ap)) {
1889 				struct cmd cmd;
1890 				cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1891 				cmd.code = 0;
1892 				cmd.idx = 0;
1893 				ace_issue_cmd(ap->regs, &cmd);
1894 			} else {
1895 				writel(0, &((ap->regs)->RxJumboPrd));
1896 				wmb();
1897 			}
1898 
1899 			ap->jumbo = 0;
1900 			ap->rx_jumbo_skbprd = 0;
1901 			printk(KERN_INFO "%s: Jumbo ring flushed\n",
1902 			       ap->name);
1903 			clear_bit(0, &ap->jumbo_refill_busy);
1904 			break;
1905 		}
1906 		default:
1907 			printk(KERN_ERR "%s: Unhandled event 0x%02x\n",
1908 			       ap->name, ap->evt_ring[evtcsm].evt);
1909 		}
1910 		evtcsm = (evtcsm + 1) % EVT_RING_ENTRIES;
1911 	}
1912 
1913 	return evtcsm;
1914 }
1915 
1916 
1917 static void ace_rx_int(struct net_device *dev, u32 rxretprd, u32 rxretcsm)
1918 {
1919 	struct ace_private *ap = netdev_priv(dev);
1920 	u32 idx;
1921 	int mini_count = 0, std_count = 0;
1922 
1923 	idx = rxretcsm;
1924 
1925 	prefetchw(&ap->cur_rx_bufs);
1926 	prefetchw(&ap->cur_mini_bufs);
1927 
1928 	while (idx != rxretprd) {
1929 		struct ring_info *rip;
1930 		struct sk_buff *skb;
1931 		struct rx_desc *retdesc;
1932 		u32 skbidx;
1933 		int bd_flags, desc_type, mapsize;
1934 		u16 csum;
1935 
1936 
1937 		/* make sure the rx descriptor isn't read before rxretprd */
1938 		if (idx == rxretcsm)
1939 			rmb();
1940 
1941 		retdesc = &ap->rx_return_ring[idx];
1942 		skbidx = retdesc->idx;
1943 		bd_flags = retdesc->flags;
1944 		desc_type = bd_flags & (BD_FLG_JUMBO | BD_FLG_MINI);
1945 
1946 		switch(desc_type) {
1947 			/*
1948 			 * Normal frames do not have any flags set
1949 			 *
1950 			 * Mini and normal frames arrive frequently,
1951 			 * so use a local counter to avoid doing
1952 			 * atomic operations for each packet arriving.
1953 			 */
1954 		case 0:
1955 			rip = &ap->skb->rx_std_skbuff[skbidx];
1956 			mapsize = ACE_STD_BUFSIZE;
1957 			std_count++;
1958 			break;
1959 		case BD_FLG_JUMBO:
1960 			rip = &ap->skb->rx_jumbo_skbuff[skbidx];
1961 			mapsize = ACE_JUMBO_BUFSIZE;
1962 			atomic_dec(&ap->cur_jumbo_bufs);
1963 			break;
1964 		case BD_FLG_MINI:
1965 			rip = &ap->skb->rx_mini_skbuff[skbidx];
1966 			mapsize = ACE_MINI_BUFSIZE;
1967 			mini_count++;
1968 			break;
1969 		default:
1970 			printk(KERN_INFO "%s: unknown frame type (0x%02x) "
1971 			       "returned by NIC\n", dev->name,
1972 			       retdesc->flags);
1973 			goto error;
1974 		}
1975 
1976 		skb = rip->skb;
1977 		rip->skb = NULL;
1978 		dma_unmap_page(&ap->pdev->dev, dma_unmap_addr(rip, mapping),
1979 			       mapsize, DMA_FROM_DEVICE);
1980 		skb_put(skb, retdesc->size);
1981 
1982 		/*
1983 		 * Fly baby, fly!
1984 		 */
1985 		csum = retdesc->tcp_udp_csum;
1986 
1987 		skb->protocol = eth_type_trans(skb, dev);
1988 
1989 		/*
1990 		 * Instead of forcing the poor tigon mips cpu to calculate
1991 		 * pseudo hdr checksum, we do this ourselves.
1992 		 */
1993 		if (bd_flags & BD_FLG_TCP_UDP_SUM) {
1994 			skb->csum = htons(csum);
1995 			skb->ip_summed = CHECKSUM_COMPLETE;
1996 		} else {
1997 			skb_checksum_none_assert(skb);
1998 		}
1999 
2000 		/* send it up */
2001 		if ((bd_flags & BD_FLG_VLAN_TAG))
2002 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), retdesc->vlan);
2003 		netif_rx(skb);
2004 
2005 		dev->stats.rx_packets++;
2006 		dev->stats.rx_bytes += retdesc->size;
2007 
2008 		idx = (idx + 1) % RX_RETURN_RING_ENTRIES;
2009 	}
2010 
2011 	atomic_sub(std_count, &ap->cur_rx_bufs);
2012 	if (!ACE_IS_TIGON_I(ap))
2013 		atomic_sub(mini_count, &ap->cur_mini_bufs);
2014 
2015  out:
2016 	/*
2017 	 * According to the documentation RxRetCsm is obsolete with
2018 	 * the 12.3.x Firmware - my Tigon I NICs seem to disagree!
2019 	 */
2020 	if (ACE_IS_TIGON_I(ap)) {
2021 		writel(idx, &ap->regs->RxRetCsm);
2022 	}
2023 	ap->cur_rx = idx;
2024 
2025 	return;
2026  error:
2027 	idx = rxretprd;
2028 	goto out;
2029 }
2030 
2031 
2032 static inline void ace_tx_int(struct net_device *dev,
2033 			      u32 txcsm, u32 idx)
2034 {
2035 	struct ace_private *ap = netdev_priv(dev);
2036 
2037 	do {
2038 		struct sk_buff *skb;
2039 		struct tx_ring_info *info;
2040 
2041 		info = ap->skb->tx_skbuff + idx;
2042 		skb = info->skb;
2043 
2044 		if (dma_unmap_len(info, maplen)) {
2045 			dma_unmap_page(&ap->pdev->dev,
2046 				       dma_unmap_addr(info, mapping),
2047 				       dma_unmap_len(info, maplen),
2048 				       DMA_TO_DEVICE);
2049 			dma_unmap_len_set(info, maplen, 0);
2050 		}
2051 
2052 		if (skb) {
2053 			dev->stats.tx_packets++;
2054 			dev->stats.tx_bytes += skb->len;
2055 			dev_consume_skb_irq(skb);
2056 			info->skb = NULL;
2057 		}
2058 
2059 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2060 	} while (idx != txcsm);
2061 
2062 	if (netif_queue_stopped(dev))
2063 		netif_wake_queue(dev);
2064 
2065 	wmb();
2066 	ap->tx_ret_csm = txcsm;
2067 
2068 	/* So... tx_ret_csm is advanced _after_ check for device wakeup.
2069 	 *
2070 	 * We could try to make it before. In this case we would get
2071 	 * the following race condition: hard_start_xmit on other cpu
2072 	 * enters after we advanced tx_ret_csm and fills space,
2073 	 * which we have just freed, so that we make illegal device wakeup.
2074 	 * There is no good way to workaround this (at entry
2075 	 * to ace_start_xmit detects this condition and prevents
2076 	 * ring corruption, but it is not a good workaround.)
2077 	 *
2078 	 * When tx_ret_csm is advanced after, we wake up device _only_
2079 	 * if we really have some space in ring (though the core doing
2080 	 * hard_start_xmit can see full ring for some period and has to
2081 	 * synchronize.) Superb.
2082 	 * BUT! We get another subtle race condition. hard_start_xmit
2083 	 * may think that ring is full between wakeup and advancing
2084 	 * tx_ret_csm and will stop device instantly! It is not so bad.
2085 	 * We are guaranteed that there is something in ring, so that
2086 	 * the next irq will resume transmission. To speedup this we could
2087 	 * mark descriptor, which closes ring with BD_FLG_COAL_NOW
2088 	 * (see ace_start_xmit).
2089 	 *
2090 	 * Well, this dilemma exists in all lock-free devices.
2091 	 * We, following scheme used in drivers by Donald Becker,
2092 	 * select the least dangerous.
2093 	 *							--ANK
2094 	 */
2095 }
2096 
2097 
2098 static irqreturn_t ace_interrupt(int irq, void *dev_id)
2099 {
2100 	struct net_device *dev = (struct net_device *)dev_id;
2101 	struct ace_private *ap = netdev_priv(dev);
2102 	struct ace_regs __iomem *regs = ap->regs;
2103 	u32 idx;
2104 	u32 txcsm, rxretcsm, rxretprd;
2105 	u32 evtcsm, evtprd;
2106 
2107 	/*
2108 	 * In case of PCI shared interrupts or spurious interrupts,
2109 	 * we want to make sure it is actually our interrupt before
2110 	 * spending any time in here.
2111 	 */
2112 	if (!(readl(&regs->HostCtrl) & IN_INT))
2113 		return IRQ_NONE;
2114 
2115 	/*
2116 	 * ACK intr now. Otherwise we will lose updates to rx_ret_prd,
2117 	 * which happened _after_ rxretprd = *ap->rx_ret_prd; but before
2118 	 * writel(0, &regs->Mb0Lo).
2119 	 *
2120 	 * "IRQ avoidance" recommended in docs applies to IRQs served
2121 	 * threads and it is wrong even for that case.
2122 	 */
2123 	writel(0, &regs->Mb0Lo);
2124 	readl(&regs->Mb0Lo);
2125 
2126 	/*
2127 	 * There is no conflict between transmit handling in
2128 	 * start_xmit and receive processing, thus there is no reason
2129 	 * to take a spin lock for RX handling. Wait until we start
2130 	 * working on the other stuff - hey we don't need a spin lock
2131 	 * anymore.
2132 	 */
2133 	rxretprd = *ap->rx_ret_prd;
2134 	rxretcsm = ap->cur_rx;
2135 
2136 	if (rxretprd != rxretcsm)
2137 		ace_rx_int(dev, rxretprd, rxretcsm);
2138 
2139 	txcsm = *ap->tx_csm;
2140 	idx = ap->tx_ret_csm;
2141 
2142 	if (txcsm != idx) {
2143 		/*
2144 		 * If each skb takes only one descriptor this check degenerates
2145 		 * to identity, because new space has just been opened.
2146 		 * But if skbs are fragmented we must check that this index
2147 		 * update releases enough of space, otherwise we just
2148 		 * wait for device to make more work.
2149 		 */
2150 		if (!tx_ring_full(ap, txcsm, ap->tx_prd))
2151 			ace_tx_int(dev, txcsm, idx);
2152 	}
2153 
2154 	evtcsm = readl(&regs->EvtCsm);
2155 	evtprd = *ap->evt_prd;
2156 
2157 	if (evtcsm != evtprd) {
2158 		evtcsm = ace_handle_event(dev, evtcsm, evtprd);
2159 		writel(evtcsm, &regs->EvtCsm);
2160 	}
2161 
2162 	/*
2163 	 * This has to go last in the interrupt handler and run with
2164 	 * the spin lock released ... what lock?
2165 	 */
2166 	if (netif_running(dev)) {
2167 		int cur_size;
2168 		int run_tasklet = 0;
2169 
2170 		cur_size = atomic_read(&ap->cur_rx_bufs);
2171 		if (cur_size < RX_LOW_STD_THRES) {
2172 			if ((cur_size < RX_PANIC_STD_THRES) &&
2173 			    !test_and_set_bit(0, &ap->std_refill_busy)) {
2174 #ifdef DEBUG
2175 				printk("low on std buffers %i\n", cur_size);
2176 #endif
2177 				ace_load_std_rx_ring(dev,
2178 						     RX_RING_SIZE - cur_size);
2179 			} else
2180 				run_tasklet = 1;
2181 		}
2182 
2183 		if (!ACE_IS_TIGON_I(ap)) {
2184 			cur_size = atomic_read(&ap->cur_mini_bufs);
2185 			if (cur_size < RX_LOW_MINI_THRES) {
2186 				if ((cur_size < RX_PANIC_MINI_THRES) &&
2187 				    !test_and_set_bit(0,
2188 						      &ap->mini_refill_busy)) {
2189 #ifdef DEBUG
2190 					printk("low on mini buffers %i\n",
2191 					       cur_size);
2192 #endif
2193 					ace_load_mini_rx_ring(dev,
2194 							      RX_MINI_SIZE - cur_size);
2195 				} else
2196 					run_tasklet = 1;
2197 			}
2198 		}
2199 
2200 		if (ap->jumbo) {
2201 			cur_size = atomic_read(&ap->cur_jumbo_bufs);
2202 			if (cur_size < RX_LOW_JUMBO_THRES) {
2203 				if ((cur_size < RX_PANIC_JUMBO_THRES) &&
2204 				    !test_and_set_bit(0,
2205 						      &ap->jumbo_refill_busy)){
2206 #ifdef DEBUG
2207 					printk("low on jumbo buffers %i\n",
2208 					       cur_size);
2209 #endif
2210 					ace_load_jumbo_rx_ring(dev,
2211 							       RX_JUMBO_SIZE - cur_size);
2212 				} else
2213 					run_tasklet = 1;
2214 			}
2215 		}
2216 		if (run_tasklet && !ap->tasklet_pending) {
2217 			ap->tasklet_pending = 1;
2218 			tasklet_schedule(&ap->ace_tasklet);
2219 		}
2220 	}
2221 
2222 	return IRQ_HANDLED;
2223 }
2224 
2225 static int ace_open(struct net_device *dev)
2226 {
2227 	struct ace_private *ap = netdev_priv(dev);
2228 	struct ace_regs __iomem *regs = ap->regs;
2229 	struct cmd cmd;
2230 
2231 	if (!(ap->fw_running)) {
2232 		printk(KERN_WARNING "%s: Firmware not running!\n", dev->name);
2233 		return -EBUSY;
2234 	}
2235 
2236 	writel(dev->mtu + ETH_HLEN + 4, &regs->IfMtu);
2237 
2238 	cmd.evt = C_CLEAR_STATS;
2239 	cmd.code = 0;
2240 	cmd.idx = 0;
2241 	ace_issue_cmd(regs, &cmd);
2242 
2243 	cmd.evt = C_HOST_STATE;
2244 	cmd.code = C_C_STACK_UP;
2245 	cmd.idx = 0;
2246 	ace_issue_cmd(regs, &cmd);
2247 
2248 	if (ap->jumbo &&
2249 	    !test_and_set_bit(0, &ap->jumbo_refill_busy))
2250 		ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE);
2251 
2252 	if (dev->flags & IFF_PROMISC) {
2253 		cmd.evt = C_SET_PROMISC_MODE;
2254 		cmd.code = C_C_PROMISC_ENABLE;
2255 		cmd.idx = 0;
2256 		ace_issue_cmd(regs, &cmd);
2257 
2258 		ap->promisc = 1;
2259 	}else
2260 		ap->promisc = 0;
2261 	ap->mcast_all = 0;
2262 
2263 #if 0
2264 	cmd.evt = C_LNK_NEGOTIATION;
2265 	cmd.code = 0;
2266 	cmd.idx = 0;
2267 	ace_issue_cmd(regs, &cmd);
2268 #endif
2269 
2270 	netif_start_queue(dev);
2271 
2272 	/*
2273 	 * Setup the bottom half rx ring refill handler
2274 	 */
2275 	tasklet_setup(&ap->ace_tasklet, ace_tasklet);
2276 	return 0;
2277 }
2278 
2279 
2280 static int ace_close(struct net_device *dev)
2281 {
2282 	struct ace_private *ap = netdev_priv(dev);
2283 	struct ace_regs __iomem *regs = ap->regs;
2284 	struct cmd cmd;
2285 	unsigned long flags;
2286 	short i;
2287 
2288 	/*
2289 	 * Without (or before) releasing irq and stopping hardware, this
2290 	 * is an absolute non-sense, by the way. It will be reset instantly
2291 	 * by the first irq.
2292 	 */
2293 	netif_stop_queue(dev);
2294 
2295 
2296 	if (ap->promisc) {
2297 		cmd.evt = C_SET_PROMISC_MODE;
2298 		cmd.code = C_C_PROMISC_DISABLE;
2299 		cmd.idx = 0;
2300 		ace_issue_cmd(regs, &cmd);
2301 		ap->promisc = 0;
2302 	}
2303 
2304 	cmd.evt = C_HOST_STATE;
2305 	cmd.code = C_C_STACK_DOWN;
2306 	cmd.idx = 0;
2307 	ace_issue_cmd(regs, &cmd);
2308 
2309 	tasklet_kill(&ap->ace_tasklet);
2310 
2311 	/*
2312 	 * Make sure one CPU is not processing packets while
2313 	 * buffers are being released by another.
2314 	 */
2315 
2316 	local_irq_save(flags);
2317 	ace_mask_irq(dev);
2318 
2319 	for (i = 0; i < ACE_TX_RING_ENTRIES(ap); i++) {
2320 		struct sk_buff *skb;
2321 		struct tx_ring_info *info;
2322 
2323 		info = ap->skb->tx_skbuff + i;
2324 		skb = info->skb;
2325 
2326 		if (dma_unmap_len(info, maplen)) {
2327 			if (ACE_IS_TIGON_I(ap)) {
2328 				/* NB: TIGON_1 is special, tx_ring is in io space */
2329 				struct tx_desc __iomem *tx;
2330 				tx = (__force struct tx_desc __iomem *) &ap->tx_ring[i];
2331 				writel(0, &tx->addr.addrhi);
2332 				writel(0, &tx->addr.addrlo);
2333 				writel(0, &tx->flagsize);
2334 			} else
2335 				memset(ap->tx_ring + i, 0,
2336 				       sizeof(struct tx_desc));
2337 			dma_unmap_page(&ap->pdev->dev,
2338 				       dma_unmap_addr(info, mapping),
2339 				       dma_unmap_len(info, maplen),
2340 				       DMA_TO_DEVICE);
2341 			dma_unmap_len_set(info, maplen, 0);
2342 		}
2343 		if (skb) {
2344 			dev_kfree_skb(skb);
2345 			info->skb = NULL;
2346 		}
2347 	}
2348 
2349 	if (ap->jumbo) {
2350 		cmd.evt = C_RESET_JUMBO_RNG;
2351 		cmd.code = 0;
2352 		cmd.idx = 0;
2353 		ace_issue_cmd(regs, &cmd);
2354 	}
2355 
2356 	ace_unmask_irq(dev);
2357 	local_irq_restore(flags);
2358 
2359 	return 0;
2360 }
2361 
2362 
2363 static inline dma_addr_t
2364 ace_map_tx_skb(struct ace_private *ap, struct sk_buff *skb,
2365 	       struct sk_buff *tail, u32 idx)
2366 {
2367 	dma_addr_t mapping;
2368 	struct tx_ring_info *info;
2369 
2370 	mapping = dma_map_page(&ap->pdev->dev, virt_to_page(skb->data),
2371 			       offset_in_page(skb->data), skb->len,
2372 			       DMA_TO_DEVICE);
2373 
2374 	info = ap->skb->tx_skbuff + idx;
2375 	info->skb = tail;
2376 	dma_unmap_addr_set(info, mapping, mapping);
2377 	dma_unmap_len_set(info, maplen, skb->len);
2378 	return mapping;
2379 }
2380 
2381 
2382 static inline void
2383 ace_load_tx_bd(struct ace_private *ap, struct tx_desc *desc, u64 addr,
2384 	       u32 flagsize, u32 vlan_tag)
2385 {
2386 #if !USE_TX_COAL_NOW
2387 	flagsize &= ~BD_FLG_COAL_NOW;
2388 #endif
2389 
2390 	if (ACE_IS_TIGON_I(ap)) {
2391 		struct tx_desc __iomem *io = (__force struct tx_desc __iomem *) desc;
2392 		writel(addr >> 32, &io->addr.addrhi);
2393 		writel(addr & 0xffffffff, &io->addr.addrlo);
2394 		writel(flagsize, &io->flagsize);
2395 		writel(vlan_tag, &io->vlanres);
2396 	} else {
2397 		desc->addr.addrhi = addr >> 32;
2398 		desc->addr.addrlo = addr;
2399 		desc->flagsize = flagsize;
2400 		desc->vlanres = vlan_tag;
2401 	}
2402 }
2403 
2404 
2405 static netdev_tx_t ace_start_xmit(struct sk_buff *skb,
2406 				  struct net_device *dev)
2407 {
2408 	struct ace_private *ap = netdev_priv(dev);
2409 	struct ace_regs __iomem *regs = ap->regs;
2410 	struct tx_desc *desc;
2411 	u32 idx, flagsize;
2412 	unsigned long maxjiff = jiffies + 3*HZ;
2413 
2414 restart:
2415 	idx = ap->tx_prd;
2416 
2417 	if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2418 		goto overflow;
2419 
2420 	if (!skb_shinfo(skb)->nr_frags)	{
2421 		dma_addr_t mapping;
2422 		u32 vlan_tag = 0;
2423 
2424 		mapping = ace_map_tx_skb(ap, skb, skb, idx);
2425 		flagsize = (skb->len << 16) | (BD_FLG_END);
2426 		if (skb->ip_summed == CHECKSUM_PARTIAL)
2427 			flagsize |= BD_FLG_TCP_UDP_SUM;
2428 		if (skb_vlan_tag_present(skb)) {
2429 			flagsize |= BD_FLG_VLAN_TAG;
2430 			vlan_tag = skb_vlan_tag_get(skb);
2431 		}
2432 		desc = ap->tx_ring + idx;
2433 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2434 
2435 		/* Look at ace_tx_int for explanations. */
2436 		if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2437 			flagsize |= BD_FLG_COAL_NOW;
2438 
2439 		ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2440 	} else {
2441 		dma_addr_t mapping;
2442 		u32 vlan_tag = 0;
2443 		int i, len = 0;
2444 
2445 		mapping = ace_map_tx_skb(ap, skb, NULL, idx);
2446 		flagsize = (skb_headlen(skb) << 16);
2447 		if (skb->ip_summed == CHECKSUM_PARTIAL)
2448 			flagsize |= BD_FLG_TCP_UDP_SUM;
2449 		if (skb_vlan_tag_present(skb)) {
2450 			flagsize |= BD_FLG_VLAN_TAG;
2451 			vlan_tag = skb_vlan_tag_get(skb);
2452 		}
2453 
2454 		ace_load_tx_bd(ap, ap->tx_ring + idx, mapping, flagsize, vlan_tag);
2455 
2456 		idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2457 
2458 		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2459 			const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2460 			struct tx_ring_info *info;
2461 
2462 			len += skb_frag_size(frag);
2463 			info = ap->skb->tx_skbuff + idx;
2464 			desc = ap->tx_ring + idx;
2465 
2466 			mapping = skb_frag_dma_map(&ap->pdev->dev, frag, 0,
2467 						   skb_frag_size(frag),
2468 						   DMA_TO_DEVICE);
2469 
2470 			flagsize = skb_frag_size(frag) << 16;
2471 			if (skb->ip_summed == CHECKSUM_PARTIAL)
2472 				flagsize |= BD_FLG_TCP_UDP_SUM;
2473 			idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2474 
2475 			if (i == skb_shinfo(skb)->nr_frags - 1) {
2476 				flagsize |= BD_FLG_END;
2477 				if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2478 					flagsize |= BD_FLG_COAL_NOW;
2479 
2480 				/*
2481 				 * Only the last fragment frees
2482 				 * the skb!
2483 				 */
2484 				info->skb = skb;
2485 			} else {
2486 				info->skb = NULL;
2487 			}
2488 			dma_unmap_addr_set(info, mapping, mapping);
2489 			dma_unmap_len_set(info, maplen, skb_frag_size(frag));
2490 			ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2491 		}
2492 	}
2493 
2494 	wmb();
2495 	ap->tx_prd = idx;
2496 	ace_set_txprd(regs, ap, idx);
2497 
2498 	if (flagsize & BD_FLG_COAL_NOW) {
2499 		netif_stop_queue(dev);
2500 
2501 		/*
2502 		 * A TX-descriptor producer (an IRQ) might have gotten
2503 		 * between, making the ring free again. Since xmit is
2504 		 * serialized, this is the only situation we have to
2505 		 * re-test.
2506 		 */
2507 		if (!tx_ring_full(ap, ap->tx_ret_csm, idx))
2508 			netif_wake_queue(dev);
2509 	}
2510 
2511 	return NETDEV_TX_OK;
2512 
2513 overflow:
2514 	/*
2515 	 * This race condition is unavoidable with lock-free drivers.
2516 	 * We wake up the queue _before_ tx_prd is advanced, so that we can
2517 	 * enter hard_start_xmit too early, while tx ring still looks closed.
2518 	 * This happens ~1-4 times per 100000 packets, so that we can allow
2519 	 * to loop syncing to other CPU. Probably, we need an additional
2520 	 * wmb() in ace_tx_intr as well.
2521 	 *
2522 	 * Note that this race is relieved by reserving one more entry
2523 	 * in tx ring than it is necessary (see original non-SG driver).
2524 	 * However, with SG we need to reserve 2*MAX_SKB_FRAGS+1, which
2525 	 * is already overkill.
2526 	 *
2527 	 * Alternative is to return with 1 not throttling queue. In this
2528 	 * case loop becomes longer, no more useful effects.
2529 	 */
2530 	if (time_before(jiffies, maxjiff)) {
2531 		barrier();
2532 		cpu_relax();
2533 		goto restart;
2534 	}
2535 
2536 	/* The ring is stuck full. */
2537 	printk(KERN_WARNING "%s: Transmit ring stuck full\n", dev->name);
2538 	return NETDEV_TX_BUSY;
2539 }
2540 
2541 
2542 static int ace_change_mtu(struct net_device *dev, int new_mtu)
2543 {
2544 	struct ace_private *ap = netdev_priv(dev);
2545 	struct ace_regs __iomem *regs = ap->regs;
2546 
2547 	writel(new_mtu + ETH_HLEN + 4, &regs->IfMtu);
2548 	dev->mtu = new_mtu;
2549 
2550 	if (new_mtu > ACE_STD_MTU) {
2551 		if (!(ap->jumbo)) {
2552 			printk(KERN_INFO "%s: Enabling Jumbo frame "
2553 			       "support\n", dev->name);
2554 			ap->jumbo = 1;
2555 			if (!test_and_set_bit(0, &ap->jumbo_refill_busy))
2556 				ace_load_jumbo_rx_ring(dev, RX_JUMBO_SIZE);
2557 			ace_set_rxtx_parms(dev, 1);
2558 		}
2559 	} else {
2560 		while (test_and_set_bit(0, &ap->jumbo_refill_busy));
2561 		ace_sync_irq(dev->irq);
2562 		ace_set_rxtx_parms(dev, 0);
2563 		if (ap->jumbo) {
2564 			struct cmd cmd;
2565 
2566 			cmd.evt = C_RESET_JUMBO_RNG;
2567 			cmd.code = 0;
2568 			cmd.idx = 0;
2569 			ace_issue_cmd(regs, &cmd);
2570 		}
2571 	}
2572 
2573 	return 0;
2574 }
2575 
2576 static int ace_get_link_ksettings(struct net_device *dev,
2577 				  struct ethtool_link_ksettings *cmd)
2578 {
2579 	struct ace_private *ap = netdev_priv(dev);
2580 	struct ace_regs __iomem *regs = ap->regs;
2581 	u32 link;
2582 	u32 supported;
2583 
2584 	memset(cmd, 0, sizeof(struct ethtool_link_ksettings));
2585 
2586 	supported = (SUPPORTED_10baseT_Half | SUPPORTED_10baseT_Full |
2587 		     SUPPORTED_100baseT_Half | SUPPORTED_100baseT_Full |
2588 		     SUPPORTED_1000baseT_Half | SUPPORTED_1000baseT_Full |
2589 		     SUPPORTED_Autoneg | SUPPORTED_FIBRE);
2590 
2591 	cmd->base.port = PORT_FIBRE;
2592 
2593 	link = readl(&regs->GigLnkState);
2594 	if (link & LNK_1000MB) {
2595 		cmd->base.speed = SPEED_1000;
2596 	} else {
2597 		link = readl(&regs->FastLnkState);
2598 		if (link & LNK_100MB)
2599 			cmd->base.speed = SPEED_100;
2600 		else if (link & LNK_10MB)
2601 			cmd->base.speed = SPEED_10;
2602 		else
2603 			cmd->base.speed = 0;
2604 	}
2605 	if (link & LNK_FULL_DUPLEX)
2606 		cmd->base.duplex = DUPLEX_FULL;
2607 	else
2608 		cmd->base.duplex = DUPLEX_HALF;
2609 
2610 	if (link & LNK_NEGOTIATE)
2611 		cmd->base.autoneg = AUTONEG_ENABLE;
2612 	else
2613 		cmd->base.autoneg = AUTONEG_DISABLE;
2614 
2615 #if 0
2616 	/*
2617 	 * Current struct ethtool_cmd is insufficient
2618 	 */
2619 	ecmd->trace = readl(&regs->TuneTrace);
2620 
2621 	ecmd->txcoal = readl(&regs->TuneTxCoalTicks);
2622 	ecmd->rxcoal = readl(&regs->TuneRxCoalTicks);
2623 #endif
2624 
2625 	ethtool_convert_legacy_u32_to_link_mode(cmd->link_modes.supported,
2626 						supported);
2627 
2628 	return 0;
2629 }
2630 
2631 static int ace_set_link_ksettings(struct net_device *dev,
2632 				  const struct ethtool_link_ksettings *cmd)
2633 {
2634 	struct ace_private *ap = netdev_priv(dev);
2635 	struct ace_regs __iomem *regs = ap->regs;
2636 	u32 link, speed;
2637 
2638 	link = readl(&regs->GigLnkState);
2639 	if (link & LNK_1000MB)
2640 		speed = SPEED_1000;
2641 	else {
2642 		link = readl(&regs->FastLnkState);
2643 		if (link & LNK_100MB)
2644 			speed = SPEED_100;
2645 		else if (link & LNK_10MB)
2646 			speed = SPEED_10;
2647 		else
2648 			speed = SPEED_100;
2649 	}
2650 
2651 	link = LNK_ENABLE | LNK_1000MB | LNK_100MB | LNK_10MB |
2652 		LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL;
2653 	if (!ACE_IS_TIGON_I(ap))
2654 		link |= LNK_TX_FLOW_CTL_Y;
2655 	if (cmd->base.autoneg == AUTONEG_ENABLE)
2656 		link |= LNK_NEGOTIATE;
2657 	if (cmd->base.speed != speed) {
2658 		link &= ~(LNK_1000MB | LNK_100MB | LNK_10MB);
2659 		switch (cmd->base.speed) {
2660 		case SPEED_1000:
2661 			link |= LNK_1000MB;
2662 			break;
2663 		case SPEED_100:
2664 			link |= LNK_100MB;
2665 			break;
2666 		case SPEED_10:
2667 			link |= LNK_10MB;
2668 			break;
2669 		}
2670 	}
2671 
2672 	if (cmd->base.duplex == DUPLEX_FULL)
2673 		link |= LNK_FULL_DUPLEX;
2674 
2675 	if (link != ap->link) {
2676 		struct cmd cmd;
2677 		printk(KERN_INFO "%s: Renegotiating link state\n",
2678 		       dev->name);
2679 
2680 		ap->link = link;
2681 		writel(link, &regs->TuneLink);
2682 		if (!ACE_IS_TIGON_I(ap))
2683 			writel(link, &regs->TuneFastLink);
2684 		wmb();
2685 
2686 		cmd.evt = C_LNK_NEGOTIATION;
2687 		cmd.code = 0;
2688 		cmd.idx = 0;
2689 		ace_issue_cmd(regs, &cmd);
2690 	}
2691 	return 0;
2692 }
2693 
2694 static void ace_get_drvinfo(struct net_device *dev,
2695 			    struct ethtool_drvinfo *info)
2696 {
2697 	struct ace_private *ap = netdev_priv(dev);
2698 
2699 	strlcpy(info->driver, "acenic", sizeof(info->driver));
2700 	snprintf(info->fw_version, sizeof(info->version), "%i.%i.%i",
2701 		 ap->firmware_major, ap->firmware_minor, ap->firmware_fix);
2702 
2703 	if (ap->pdev)
2704 		strlcpy(info->bus_info, pci_name(ap->pdev),
2705 			sizeof(info->bus_info));
2706 
2707 }
2708 
2709 /*
2710  * Set the hardware MAC address.
2711  */
2712 static int ace_set_mac_addr(struct net_device *dev, void *p)
2713 {
2714 	struct ace_private *ap = netdev_priv(dev);
2715 	struct ace_regs __iomem *regs = ap->regs;
2716 	struct sockaddr *addr=p;
2717 	const u8 *da;
2718 	struct cmd cmd;
2719 
2720 	if(netif_running(dev))
2721 		return -EBUSY;
2722 
2723 	eth_hw_addr_set(dev, addr->sa_data);
2724 
2725 	da = (const u8 *)dev->dev_addr;
2726 
2727 	writel(da[0] << 8 | da[1], &regs->MacAddrHi);
2728 	writel((da[2] << 24) | (da[3] << 16) | (da[4] << 8) | da[5],
2729 	       &regs->MacAddrLo);
2730 
2731 	cmd.evt = C_SET_MAC_ADDR;
2732 	cmd.code = 0;
2733 	cmd.idx = 0;
2734 	ace_issue_cmd(regs, &cmd);
2735 
2736 	return 0;
2737 }
2738 
2739 
2740 static void ace_set_multicast_list(struct net_device *dev)
2741 {
2742 	struct ace_private *ap = netdev_priv(dev);
2743 	struct ace_regs __iomem *regs = ap->regs;
2744 	struct cmd cmd;
2745 
2746 	if ((dev->flags & IFF_ALLMULTI) && !(ap->mcast_all)) {
2747 		cmd.evt = C_SET_MULTICAST_MODE;
2748 		cmd.code = C_C_MCAST_ENABLE;
2749 		cmd.idx = 0;
2750 		ace_issue_cmd(regs, &cmd);
2751 		ap->mcast_all = 1;
2752 	} else if (ap->mcast_all) {
2753 		cmd.evt = C_SET_MULTICAST_MODE;
2754 		cmd.code = C_C_MCAST_DISABLE;
2755 		cmd.idx = 0;
2756 		ace_issue_cmd(regs, &cmd);
2757 		ap->mcast_all = 0;
2758 	}
2759 
2760 	if ((dev->flags & IFF_PROMISC) && !(ap->promisc)) {
2761 		cmd.evt = C_SET_PROMISC_MODE;
2762 		cmd.code = C_C_PROMISC_ENABLE;
2763 		cmd.idx = 0;
2764 		ace_issue_cmd(regs, &cmd);
2765 		ap->promisc = 1;
2766 	}else if (!(dev->flags & IFF_PROMISC) && (ap->promisc)) {
2767 		cmd.evt = C_SET_PROMISC_MODE;
2768 		cmd.code = C_C_PROMISC_DISABLE;
2769 		cmd.idx = 0;
2770 		ace_issue_cmd(regs, &cmd);
2771 		ap->promisc = 0;
2772 	}
2773 
2774 	/*
2775 	 * For the time being multicast relies on the upper layers
2776 	 * filtering it properly. The Firmware does not allow one to
2777 	 * set the entire multicast list at a time and keeping track of
2778 	 * it here is going to be messy.
2779 	 */
2780 	if (!netdev_mc_empty(dev) && !ap->mcast_all) {
2781 		cmd.evt = C_SET_MULTICAST_MODE;
2782 		cmd.code = C_C_MCAST_ENABLE;
2783 		cmd.idx = 0;
2784 		ace_issue_cmd(regs, &cmd);
2785 	}else if (!ap->mcast_all) {
2786 		cmd.evt = C_SET_MULTICAST_MODE;
2787 		cmd.code = C_C_MCAST_DISABLE;
2788 		cmd.idx = 0;
2789 		ace_issue_cmd(regs, &cmd);
2790 	}
2791 }
2792 
2793 
2794 static struct net_device_stats *ace_get_stats(struct net_device *dev)
2795 {
2796 	struct ace_private *ap = netdev_priv(dev);
2797 	struct ace_mac_stats __iomem *mac_stats =
2798 		(struct ace_mac_stats __iomem *)ap->regs->Stats;
2799 
2800 	dev->stats.rx_missed_errors = readl(&mac_stats->drop_space);
2801 	dev->stats.multicast = readl(&mac_stats->kept_mc);
2802 	dev->stats.collisions = readl(&mac_stats->coll);
2803 
2804 	return &dev->stats;
2805 }
2806 
2807 
2808 static void ace_copy(struct ace_regs __iomem *regs, const __be32 *src,
2809 		     u32 dest, int size)
2810 {
2811 	void __iomem *tdest;
2812 	short tsize, i;
2813 
2814 	if (size <= 0)
2815 		return;
2816 
2817 	while (size > 0) {
2818 		tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2819 			    min_t(u32, size, ACE_WINDOW_SIZE));
2820 		tdest = (void __iomem *) &regs->Window +
2821 			(dest & (ACE_WINDOW_SIZE - 1));
2822 		writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2823 		for (i = 0; i < (tsize / 4); i++) {
2824 			/* Firmware is big-endian */
2825 			writel(be32_to_cpup(src), tdest);
2826 			src++;
2827 			tdest += 4;
2828 			dest += 4;
2829 			size -= 4;
2830 		}
2831 	}
2832 }
2833 
2834 
2835 static void ace_clear(struct ace_regs __iomem *regs, u32 dest, int size)
2836 {
2837 	void __iomem *tdest;
2838 	short tsize = 0, i;
2839 
2840 	if (size <= 0)
2841 		return;
2842 
2843 	while (size > 0) {
2844 		tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2845 				min_t(u32, size, ACE_WINDOW_SIZE));
2846 		tdest = (void __iomem *) &regs->Window +
2847 			(dest & (ACE_WINDOW_SIZE - 1));
2848 		writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2849 
2850 		for (i = 0; i < (tsize / 4); i++) {
2851 			writel(0, tdest + i*4);
2852 		}
2853 
2854 		dest += tsize;
2855 		size -= tsize;
2856 	}
2857 }
2858 
2859 
2860 /*
2861  * Download the firmware into the SRAM on the NIC
2862  *
2863  * This operation requires the NIC to be halted and is performed with
2864  * interrupts disabled and with the spinlock hold.
2865  */
2866 static int ace_load_firmware(struct net_device *dev)
2867 {
2868 	const struct firmware *fw;
2869 	const char *fw_name = "acenic/tg2.bin";
2870 	struct ace_private *ap = netdev_priv(dev);
2871 	struct ace_regs __iomem *regs = ap->regs;
2872 	const __be32 *fw_data;
2873 	u32 load_addr;
2874 	int ret;
2875 
2876 	if (!(readl(&regs->CpuCtrl) & CPU_HALTED)) {
2877 		printk(KERN_ERR "%s: trying to download firmware while the "
2878 		       "CPU is running!\n", ap->name);
2879 		return -EFAULT;
2880 	}
2881 
2882 	if (ACE_IS_TIGON_I(ap))
2883 		fw_name = "acenic/tg1.bin";
2884 
2885 	ret = request_firmware(&fw, fw_name, &ap->pdev->dev);
2886 	if (ret) {
2887 		printk(KERN_ERR "%s: Failed to load firmware \"%s\"\n",
2888 		       ap->name, fw_name);
2889 		return ret;
2890 	}
2891 
2892 	fw_data = (void *)fw->data;
2893 
2894 	/* Firmware blob starts with version numbers, followed by
2895 	   load and start address. Remainder is the blob to be loaded
2896 	   contiguously from load address. We don't bother to represent
2897 	   the BSS/SBSS sections any more, since we were clearing the
2898 	   whole thing anyway. */
2899 	ap->firmware_major = fw->data[0];
2900 	ap->firmware_minor = fw->data[1];
2901 	ap->firmware_fix = fw->data[2];
2902 
2903 	ap->firmware_start = be32_to_cpu(fw_data[1]);
2904 	if (ap->firmware_start < 0x4000 || ap->firmware_start >= 0x80000) {
2905 		printk(KERN_ERR "%s: bogus load address %08x in \"%s\"\n",
2906 		       ap->name, ap->firmware_start, fw_name);
2907 		ret = -EINVAL;
2908 		goto out;
2909 	}
2910 
2911 	load_addr = be32_to_cpu(fw_data[2]);
2912 	if (load_addr < 0x4000 || load_addr >= 0x80000) {
2913 		printk(KERN_ERR "%s: bogus load address %08x in \"%s\"\n",
2914 		       ap->name, load_addr, fw_name);
2915 		ret = -EINVAL;
2916 		goto out;
2917 	}
2918 
2919 	/*
2920 	 * Do not try to clear more than 512KiB or we end up seeing
2921 	 * funny things on NICs with only 512KiB SRAM
2922 	 */
2923 	ace_clear(regs, 0x2000, 0x80000-0x2000);
2924 	ace_copy(regs, &fw_data[3], load_addr, fw->size-12);
2925  out:
2926 	release_firmware(fw);
2927 	return ret;
2928 }
2929 
2930 
2931 /*
2932  * The eeprom on the AceNIC is an Atmel i2c EEPROM.
2933  *
2934  * Accessing the EEPROM is `interesting' to say the least - don't read
2935  * this code right after dinner.
2936  *
2937  * This is all about black magic and bit-banging the device .... I
2938  * wonder in what hospital they have put the guy who designed the i2c
2939  * specs.
2940  *
2941  * Oh yes, this is only the beginning!
2942  *
2943  * Thanks to Stevarino Webinski for helping tracking down the bugs in the
2944  * code i2c readout code by beta testing all my hacks.
2945  */
2946 static void eeprom_start(struct ace_regs __iomem *regs)
2947 {
2948 	u32 local;
2949 
2950 	readl(&regs->LocalCtrl);
2951 	udelay(ACE_SHORT_DELAY);
2952 	local = readl(&regs->LocalCtrl);
2953 	local |= EEPROM_DATA_OUT | EEPROM_WRITE_ENABLE;
2954 	writel(local, &regs->LocalCtrl);
2955 	readl(&regs->LocalCtrl);
2956 	mb();
2957 	udelay(ACE_SHORT_DELAY);
2958 	local |= EEPROM_CLK_OUT;
2959 	writel(local, &regs->LocalCtrl);
2960 	readl(&regs->LocalCtrl);
2961 	mb();
2962 	udelay(ACE_SHORT_DELAY);
2963 	local &= ~EEPROM_DATA_OUT;
2964 	writel(local, &regs->LocalCtrl);
2965 	readl(&regs->LocalCtrl);
2966 	mb();
2967 	udelay(ACE_SHORT_DELAY);
2968 	local &= ~EEPROM_CLK_OUT;
2969 	writel(local, &regs->LocalCtrl);
2970 	readl(&regs->LocalCtrl);
2971 	mb();
2972 }
2973 
2974 
2975 static void eeprom_prep(struct ace_regs __iomem *regs, u8 magic)
2976 {
2977 	short i;
2978 	u32 local;
2979 
2980 	udelay(ACE_SHORT_DELAY);
2981 	local = readl(&regs->LocalCtrl);
2982 	local &= ~EEPROM_DATA_OUT;
2983 	local |= EEPROM_WRITE_ENABLE;
2984 	writel(local, &regs->LocalCtrl);
2985 	readl(&regs->LocalCtrl);
2986 	mb();
2987 
2988 	for (i = 0; i < 8; i++, magic <<= 1) {
2989 		udelay(ACE_SHORT_DELAY);
2990 		if (magic & 0x80)
2991 			local |= EEPROM_DATA_OUT;
2992 		else
2993 			local &= ~EEPROM_DATA_OUT;
2994 		writel(local, &regs->LocalCtrl);
2995 		readl(&regs->LocalCtrl);
2996 		mb();
2997 
2998 		udelay(ACE_SHORT_DELAY);
2999 		local |= EEPROM_CLK_OUT;
3000 		writel(local, &regs->LocalCtrl);
3001 		readl(&regs->LocalCtrl);
3002 		mb();
3003 		udelay(ACE_SHORT_DELAY);
3004 		local &= ~(EEPROM_CLK_OUT | EEPROM_DATA_OUT);
3005 		writel(local, &regs->LocalCtrl);
3006 		readl(&regs->LocalCtrl);
3007 		mb();
3008 	}
3009 }
3010 
3011 
3012 static int eeprom_check_ack(struct ace_regs __iomem *regs)
3013 {
3014 	int state;
3015 	u32 local;
3016 
3017 	local = readl(&regs->LocalCtrl);
3018 	local &= ~EEPROM_WRITE_ENABLE;
3019 	writel(local, &regs->LocalCtrl);
3020 	readl(&regs->LocalCtrl);
3021 	mb();
3022 	udelay(ACE_LONG_DELAY);
3023 	local |= EEPROM_CLK_OUT;
3024 	writel(local, &regs->LocalCtrl);
3025 	readl(&regs->LocalCtrl);
3026 	mb();
3027 	udelay(ACE_SHORT_DELAY);
3028 	/* sample data in middle of high clk */
3029 	state = (readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0;
3030 	udelay(ACE_SHORT_DELAY);
3031 	mb();
3032 	writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3033 	readl(&regs->LocalCtrl);
3034 	mb();
3035 
3036 	return state;
3037 }
3038 
3039 
3040 static void eeprom_stop(struct ace_regs __iomem *regs)
3041 {
3042 	u32 local;
3043 
3044 	udelay(ACE_SHORT_DELAY);
3045 	local = readl(&regs->LocalCtrl);
3046 	local |= EEPROM_WRITE_ENABLE;
3047 	writel(local, &regs->LocalCtrl);
3048 	readl(&regs->LocalCtrl);
3049 	mb();
3050 	udelay(ACE_SHORT_DELAY);
3051 	local &= ~EEPROM_DATA_OUT;
3052 	writel(local, &regs->LocalCtrl);
3053 	readl(&regs->LocalCtrl);
3054 	mb();
3055 	udelay(ACE_SHORT_DELAY);
3056 	local |= EEPROM_CLK_OUT;
3057 	writel(local, &regs->LocalCtrl);
3058 	readl(&regs->LocalCtrl);
3059 	mb();
3060 	udelay(ACE_SHORT_DELAY);
3061 	local |= EEPROM_DATA_OUT;
3062 	writel(local, &regs->LocalCtrl);
3063 	readl(&regs->LocalCtrl);
3064 	mb();
3065 	udelay(ACE_LONG_DELAY);
3066 	local &= ~EEPROM_CLK_OUT;
3067 	writel(local, &regs->LocalCtrl);
3068 	mb();
3069 }
3070 
3071 
3072 /*
3073  * Read a whole byte from the EEPROM.
3074  */
3075 static int read_eeprom_byte(struct net_device *dev, unsigned long offset)
3076 {
3077 	struct ace_private *ap = netdev_priv(dev);
3078 	struct ace_regs __iomem *regs = ap->regs;
3079 	unsigned long flags;
3080 	u32 local;
3081 	int result = 0;
3082 	short i;
3083 
3084 	/*
3085 	 * Don't take interrupts on this CPU will bit banging
3086 	 * the %#%#@$ I2C device
3087 	 */
3088 	local_irq_save(flags);
3089 
3090 	eeprom_start(regs);
3091 
3092 	eeprom_prep(regs, EEPROM_WRITE_SELECT);
3093 	if (eeprom_check_ack(regs)) {
3094 		local_irq_restore(flags);
3095 		printk(KERN_ERR "%s: Unable to sync eeprom\n", ap->name);
3096 		result = -EIO;
3097 		goto eeprom_read_error;
3098 	}
3099 
3100 	eeprom_prep(regs, (offset >> 8) & 0xff);
3101 	if (eeprom_check_ack(regs)) {
3102 		local_irq_restore(flags);
3103 		printk(KERN_ERR "%s: Unable to set address byte 0\n",
3104 		       ap->name);
3105 		result = -EIO;
3106 		goto eeprom_read_error;
3107 	}
3108 
3109 	eeprom_prep(regs, offset & 0xff);
3110 	if (eeprom_check_ack(regs)) {
3111 		local_irq_restore(flags);
3112 		printk(KERN_ERR "%s: Unable to set address byte 1\n",
3113 		       ap->name);
3114 		result = -EIO;
3115 		goto eeprom_read_error;
3116 	}
3117 
3118 	eeprom_start(regs);
3119 	eeprom_prep(regs, EEPROM_READ_SELECT);
3120 	if (eeprom_check_ack(regs)) {
3121 		local_irq_restore(flags);
3122 		printk(KERN_ERR "%s: Unable to set READ_SELECT\n",
3123 		       ap->name);
3124 		result = -EIO;
3125 		goto eeprom_read_error;
3126 	}
3127 
3128 	for (i = 0; i < 8; i++) {
3129 		local = readl(&regs->LocalCtrl);
3130 		local &= ~EEPROM_WRITE_ENABLE;
3131 		writel(local, &regs->LocalCtrl);
3132 		readl(&regs->LocalCtrl);
3133 		udelay(ACE_LONG_DELAY);
3134 		mb();
3135 		local |= EEPROM_CLK_OUT;
3136 		writel(local, &regs->LocalCtrl);
3137 		readl(&regs->LocalCtrl);
3138 		mb();
3139 		udelay(ACE_SHORT_DELAY);
3140 		/* sample data mid high clk */
3141 		result = (result << 1) |
3142 			((readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0);
3143 		udelay(ACE_SHORT_DELAY);
3144 		mb();
3145 		local = readl(&regs->LocalCtrl);
3146 		local &= ~EEPROM_CLK_OUT;
3147 		writel(local, &regs->LocalCtrl);
3148 		readl(&regs->LocalCtrl);
3149 		udelay(ACE_SHORT_DELAY);
3150 		mb();
3151 		if (i == 7) {
3152 			local |= EEPROM_WRITE_ENABLE;
3153 			writel(local, &regs->LocalCtrl);
3154 			readl(&regs->LocalCtrl);
3155 			mb();
3156 			udelay(ACE_SHORT_DELAY);
3157 		}
3158 	}
3159 
3160 	local |= EEPROM_DATA_OUT;
3161 	writel(local, &regs->LocalCtrl);
3162 	readl(&regs->LocalCtrl);
3163 	mb();
3164 	udelay(ACE_SHORT_DELAY);
3165 	writel(readl(&regs->LocalCtrl) | EEPROM_CLK_OUT, &regs->LocalCtrl);
3166 	readl(&regs->LocalCtrl);
3167 	udelay(ACE_LONG_DELAY);
3168 	writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3169 	readl(&regs->LocalCtrl);
3170 	mb();
3171 	udelay(ACE_SHORT_DELAY);
3172 	eeprom_stop(regs);
3173 
3174 	local_irq_restore(flags);
3175  out:
3176 	return result;
3177 
3178  eeprom_read_error:
3179 	printk(KERN_ERR "%s: Unable to read eeprom byte 0x%02lx\n",
3180 	       ap->name, offset);
3181 	goto out;
3182 }
3183 
3184 module_pci_driver(acenic_pci_driver);
3185