]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - drivers/net/acenic.c
[NET]: Replace CHECKSUM_HW by CHECKSUM_PARTIAL/CHECKSUM_COMPLETE
[mirror_ubuntu-artful-kernel.git] / drivers / net / acenic.c
1 /*
2 * acenic.c: Linux driver for the Alteon AceNIC Gigabit Ethernet card
3 * and other Tigon based cards.
4 *
5 * Copyright 1998-2002 by Jes Sorensen, <jes@trained-monkey.org>.
6 *
7 * Thanks to Alteon and 3Com for providing hardware and documentation
8 * enabling me to write this driver.
9 *
10 * A mailing list for discussing the use of this driver has been
11 * setup, please subscribe to the lists if you have any questions
12 * about the driver. Send mail to linux-acenic-help@sunsite.auc.dk to
13 * see how to subscribe.
14 *
15 * This program is free software; you can redistribute it and/or modify
16 * it under the terms of the GNU General Public License as published by
17 * the Free Software Foundation; either version 2 of the License, or
18 * (at your option) any later version.
19 *
20 * Additional credits:
21 * Pete Wyckoff <wyckoff@ca.sandia.gov>: Initial Linux/Alpha and trace
22 * dump support. The trace dump support has not been
23 * integrated yet however.
24 * Troy Benjegerdes: Big Endian (PPC) patches.
25 * Nate Stahl: Better out of memory handling and stats support.
26 * Aman Singla: Nasty race between interrupt handler and tx code dealing
27 * with 'testing the tx_ret_csm and setting tx_full'
28 * David S. Miller <davem@redhat.com>: conversion to new PCI dma mapping
29 * infrastructure and Sparc support
30 * Pierrick Pinasseau (CERN): For lending me an Ultra 5 to test the
31 * driver under Linux/Sparc64
32 * Matt Domsch <Matt_Domsch@dell.com>: Detect Alteon 1000baseT cards
33 * ETHTOOL_GDRVINFO support
34 * Chip Salzenberg <chip@valinux.com>: Fix race condition between tx
35 * handler and close() cleanup.
36 * Ken Aaker <kdaaker@rchland.vnet.ibm.com>: Correct check for whether
37 * memory mapped IO is enabled to
38 * make the driver work on RS/6000.
39 * Takayoshi Kouchi <kouchi@hpc.bs1.fc.nec.co.jp>: Identifying problem
40 * where the driver would disable
41 * bus master mode if it had to disable
42 * write and invalidate.
43 * Stephen Hack <stephen_hack@hp.com>: Fixed ace_set_mac_addr for little
44 * endian systems.
45 * Val Henson <vhenson@esscom.com>: Reset Jumbo skb producer and
46 * rx producer index when
47 * flushing the Jumbo ring.
48 * Hans Grobler <grobh@sun.ac.za>: Memory leak fixes in the
49 * driver init path.
50 * Grant Grundler <grundler@cup.hp.com>: PCI write posting fixes.
51 */
52
53 #include <linux/module.h>
54 #include <linux/moduleparam.h>
55 #include <linux/version.h>
56 #include <linux/types.h>
57 #include <linux/errno.h>
58 #include <linux/ioport.h>
59 #include <linux/pci.h>
60 #include <linux/dma-mapping.h>
61 #include <linux/kernel.h>
62 #include <linux/netdevice.h>
63 #include <linux/etherdevice.h>
64 #include <linux/skbuff.h>
65 #include <linux/init.h>
66 #include <linux/delay.h>
67 #include <linux/mm.h>
68 #include <linux/highmem.h>
69 #include <linux/sockios.h>
70
71 #if defined(CONFIG_VLAN_8021Q) || defined(CONFIG_VLAN_8021Q_MODULE)
72 #include <linux/if_vlan.h>
73 #endif
74
75 #ifdef SIOCETHTOOL
76 #include <linux/ethtool.h>
77 #endif
78
79 #include <net/sock.h>
80 #include <net/ip.h>
81
82 #include <asm/system.h>
83 #include <asm/io.h>
84 #include <asm/irq.h>
85 #include <asm/byteorder.h>
86 #include <asm/uaccess.h>
87
88
89 #define DRV_NAME "acenic"
90
91 #undef INDEX_DEBUG
92
93 #ifdef CONFIG_ACENIC_OMIT_TIGON_I
94 #define ACE_IS_TIGON_I(ap) 0
95 #define ACE_TX_RING_ENTRIES(ap) MAX_TX_RING_ENTRIES
96 #else
97 #define ACE_IS_TIGON_I(ap) (ap->version == 1)
98 #define ACE_TX_RING_ENTRIES(ap) ap->tx_ring_entries
99 #endif
100
101 #ifndef PCI_VENDOR_ID_ALTEON
102 #define PCI_VENDOR_ID_ALTEON 0x12ae
103 #endif
104 #ifndef PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE
105 #define PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE 0x0001
106 #define PCI_DEVICE_ID_ALTEON_ACENIC_COPPER 0x0002
107 #endif
108 #ifndef PCI_DEVICE_ID_3COM_3C985
109 #define PCI_DEVICE_ID_3COM_3C985 0x0001
110 #endif
111 #ifndef PCI_VENDOR_ID_NETGEAR
112 #define PCI_VENDOR_ID_NETGEAR 0x1385
113 #define PCI_DEVICE_ID_NETGEAR_GA620 0x620a
114 #endif
115 #ifndef PCI_DEVICE_ID_NETGEAR_GA620T
116 #define PCI_DEVICE_ID_NETGEAR_GA620T 0x630a
117 #endif
118
119
120 /*
121 * Farallon used the DEC vendor ID by mistake and they seem not
122 * to care - stinky!
123 */
124 #ifndef PCI_DEVICE_ID_FARALLON_PN9000SX
125 #define PCI_DEVICE_ID_FARALLON_PN9000SX 0x1a
126 #endif
127 #ifndef PCI_DEVICE_ID_FARALLON_PN9100T
128 #define PCI_DEVICE_ID_FARALLON_PN9100T 0xfa
129 #endif
130 #ifndef PCI_VENDOR_ID_SGI
131 #define PCI_VENDOR_ID_SGI 0x10a9
132 #endif
133 #ifndef PCI_DEVICE_ID_SGI_ACENIC
134 #define PCI_DEVICE_ID_SGI_ACENIC 0x0009
135 #endif
136
137 static struct pci_device_id acenic_pci_tbl[] = {
138 { PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE,
139 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
140 { PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_ALTEON_ACENIC_COPPER,
141 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
142 { PCI_VENDOR_ID_3COM, PCI_DEVICE_ID_3COM_3C985,
143 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
144 { PCI_VENDOR_ID_NETGEAR, PCI_DEVICE_ID_NETGEAR_GA620,
145 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
146 { PCI_VENDOR_ID_NETGEAR, PCI_DEVICE_ID_NETGEAR_GA620T,
147 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
148 /*
149 * Farallon used the DEC vendor ID on their cards incorrectly,
150 * then later Alteon's ID.
151 */
152 { PCI_VENDOR_ID_DEC, PCI_DEVICE_ID_FARALLON_PN9000SX,
153 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
154 { PCI_VENDOR_ID_ALTEON, PCI_DEVICE_ID_FARALLON_PN9100T,
155 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
156 { PCI_VENDOR_ID_SGI, PCI_DEVICE_ID_SGI_ACENIC,
157 PCI_ANY_ID, PCI_ANY_ID, PCI_CLASS_NETWORK_ETHERNET << 8, 0xffff00, },
158 { }
159 };
160 MODULE_DEVICE_TABLE(pci, acenic_pci_tbl);
161
162 #ifndef SET_NETDEV_DEV
163 #define SET_NETDEV_DEV(net, pdev) do{} while(0)
164 #endif
165
166 #if LINUX_VERSION_CODE >= 0x2051c
167 #define ace_sync_irq(irq) synchronize_irq(irq)
168 #else
169 #define ace_sync_irq(irq) synchronize_irq()
170 #endif
171
172 #ifndef offset_in_page
173 #define offset_in_page(ptr) ((unsigned long)(ptr) & ~PAGE_MASK)
174 #endif
175
176 #define ACE_MAX_MOD_PARMS 8
177 #define BOARD_IDX_STATIC 0
178 #define BOARD_IDX_OVERFLOW -1
179
180 #if (defined(CONFIG_VLAN_8021Q) || defined(CONFIG_VLAN_8021Q_MODULE)) && \
181 defined(NETIF_F_HW_VLAN_RX)
182 #define ACENIC_DO_VLAN 1
183 #define ACE_RCB_VLAN_FLAG RCB_FLG_VLAN_ASSIST
184 #else
185 #define ACENIC_DO_VLAN 0
186 #define ACE_RCB_VLAN_FLAG 0
187 #endif
188
189 #include "acenic.h"
190
191 /*
192 * These must be defined before the firmware is included.
193 */
194 #define MAX_TEXT_LEN 96*1024
195 #define MAX_RODATA_LEN 8*1024
196 #define MAX_DATA_LEN 2*1024
197
198 #include "acenic_firmware.h"
199
200 #ifndef tigon2FwReleaseLocal
201 #define tigon2FwReleaseLocal 0
202 #endif
203
204 /*
205 * This driver currently supports Tigon I and Tigon II based cards
206 * including the Alteon AceNIC, the 3Com 3C985[B] and NetGear
207 * GA620. The driver should also work on the SGI, DEC and Farallon
208 * versions of the card, however I have not been able to test that
209 * myself.
210 *
211 * This card is really neat, it supports receive hardware checksumming
212 * and jumbo frames (up to 9000 bytes) and does a lot of work in the
213 * firmware. Also the programming interface is quite neat, except for
214 * the parts dealing with the i2c eeprom on the card ;-)
215 *
216 * Using jumbo frames:
217 *
218 * To enable jumbo frames, simply specify an mtu between 1500 and 9000
219 * bytes to ifconfig. Jumbo frames can be enabled or disabled at any time
220 * by running `ifconfig eth<X> mtu <MTU>' with <X> being the Ethernet
221 * interface number and <MTU> being the MTU value.
222 *
223 * Module parameters:
224 *
225 * When compiled as a loadable module, the driver allows for a number
226 * of module parameters to be specified. The driver supports the
227 * following module parameters:
228 *
229 * trace=<val> - Firmware trace level. This requires special traced
230 * firmware to replace the firmware supplied with
231 * the driver - for debugging purposes only.
232 *
233 * link=<val> - Link state. Normally you want to use the default link
234 * parameters set by the driver. This can be used to
235 * override these in case your switch doesn't negotiate
236 * the link properly. Valid values are:
237 * 0x0001 - Force half duplex link.
238 * 0x0002 - Do not negotiate line speed with the other end.
239 * 0x0010 - 10Mbit/sec link.
240 * 0x0020 - 100Mbit/sec link.
241 * 0x0040 - 1000Mbit/sec link.
242 * 0x0100 - Do not negotiate flow control.
243 * 0x0200 - Enable RX flow control Y
244 * 0x0400 - Enable TX flow control Y (Tigon II NICs only).
245 * Default value is 0x0270, ie. enable link+flow
246 * control negotiation. Negotiating the highest
247 * possible link speed with RX flow control enabled.
248 *
249 * When disabling link speed negotiation, only one link
250 * speed is allowed to be specified!
251 *
252 * tx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
253 * to wait for more packets to arive before
254 * interrupting the host, from the time the first
255 * packet arrives.
256 *
257 * rx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
258 * to wait for more packets to arive in the transmit ring,
259 * before interrupting the host, after transmitting the
260 * first packet in the ring.
261 *
262 * max_tx_desc=<val> - maximum number of transmit descriptors
263 * (packets) transmitted before interrupting the host.
264 *
265 * max_rx_desc=<val> - maximum number of receive descriptors
266 * (packets) received before interrupting the host.
267 *
268 * tx_ratio=<val> - 7 bit value (0 - 63) specifying the split in 64th
269 * increments of the NIC's on board memory to be used for
270 * transmit and receive buffers. For the 1MB NIC app. 800KB
271 * is available, on the 1/2MB NIC app. 300KB is available.
272 * 68KB will always be available as a minimum for both
273 * directions. The default value is a 50/50 split.
274 * dis_pci_mem_inval=<val> - disable PCI memory write and invalidate
275 * operations, default (1) is to always disable this as
276 * that is what Alteon does on NT. I have not been able
277 * to measure any real performance differences with
278 * this on my systems. Set <val>=0 if you want to
279 * enable these operations.
280 *
281 * If you use more than one NIC, specify the parameters for the
282 * individual NICs with a comma, ie. trace=0,0x00001fff,0 you want to
283 * run tracing on NIC #2 but not on NIC #1 and #3.
284 *
285 * TODO:
286 *
287 * - Proper multicast support.
288 * - NIC dump support.
289 * - More tuning parameters.
290 *
291 * The mini ring is not used under Linux and I am not sure it makes sense
292 * to actually use it.
293 *
294 * New interrupt handler strategy:
295 *
296 * The old interrupt handler worked using the traditional method of
297 * replacing an skbuff with a new one when a packet arrives. However
298 * the rx rings do not need to contain a static number of buffer
299 * descriptors, thus it makes sense to move the memory allocation out
300 * of the main interrupt handler and do it in a bottom half handler
301 * and only allocate new buffers when the number of buffers in the
302 * ring is below a certain threshold. In order to avoid starving the
303 * NIC under heavy load it is however necessary to force allocation
304 * when hitting a minimum threshold. The strategy for alloction is as
305 * follows:
306 *
307 * RX_LOW_BUF_THRES - allocate buffers in the bottom half
308 * RX_PANIC_LOW_THRES - we are very low on buffers, allocate
309 * the buffers in the interrupt handler
310 * RX_RING_THRES - maximum number of buffers in the rx ring
311 * RX_MINI_THRES - maximum number of buffers in the mini ring
312 * RX_JUMBO_THRES - maximum number of buffers in the jumbo ring
313 *
314 * One advantagous side effect of this allocation approach is that the
315 * entire rx processing can be done without holding any spin lock
316 * since the rx rings and registers are totally independent of the tx
317 * ring and its registers. This of course includes the kmalloc's of
318 * new skb's. Thus start_xmit can run in parallel with rx processing
319 * and the memory allocation on SMP systems.
320 *
321 * Note that running the skb reallocation in a bottom half opens up
322 * another can of races which needs to be handled properly. In
323 * particular it can happen that the interrupt handler tries to run
324 * the reallocation while the bottom half is either running on another
325 * CPU or was interrupted on the same CPU. To get around this the
326 * driver uses bitops to prevent the reallocation routines from being
327 * reentered.
328 *
329 * TX handling can also be done without holding any spin lock, wheee
330 * this is fun! since tx_ret_csm is only written to by the interrupt
331 * handler. The case to be aware of is when shutting down the device
332 * and cleaning up where it is necessary to make sure that
333 * start_xmit() is not running while this is happening. Well DaveM
334 * informs me that this case is already protected against ... bye bye
335 * Mr. Spin Lock, it was nice to know you.
336 *
337 * TX interrupts are now partly disabled so the NIC will only generate
338 * TX interrupts for the number of coal ticks, not for the number of
339 * TX packets in the queue. This should reduce the number of TX only,
340 * ie. when no RX processing is done, interrupts seen.
341 */
342
343 /*
344 * Threshold values for RX buffer allocation - the low water marks for
345 * when to start refilling the rings are set to 75% of the ring
346 * sizes. It seems to make sense to refill the rings entirely from the
347 * intrrupt handler once it gets below the panic threshold, that way
348 * we don't risk that the refilling is moved to another CPU when the
349 * one running the interrupt handler just got the slab code hot in its
350 * cache.
351 */
352 #define RX_RING_SIZE 72
353 #define RX_MINI_SIZE 64
354 #define RX_JUMBO_SIZE 48
355
356 #define RX_PANIC_STD_THRES 16
357 #define RX_PANIC_STD_REFILL (3*RX_PANIC_STD_THRES)/2
358 #define RX_LOW_STD_THRES (3*RX_RING_SIZE)/4
359 #define RX_PANIC_MINI_THRES 12
360 #define RX_PANIC_MINI_REFILL (3*RX_PANIC_MINI_THRES)/2
361 #define RX_LOW_MINI_THRES (3*RX_MINI_SIZE)/4
362 #define RX_PANIC_JUMBO_THRES 6
363 #define RX_PANIC_JUMBO_REFILL (3*RX_PANIC_JUMBO_THRES)/2
364 #define RX_LOW_JUMBO_THRES (3*RX_JUMBO_SIZE)/4
365
366
367 /*
368 * Size of the mini ring entries, basically these just should be big
369 * enough to take TCP ACKs
370 */
371 #define ACE_MINI_SIZE 100
372
373 #define ACE_MINI_BUFSIZE ACE_MINI_SIZE
374 #define ACE_STD_BUFSIZE (ACE_STD_MTU + ETH_HLEN + 4)
375 #define ACE_JUMBO_BUFSIZE (ACE_JUMBO_MTU + ETH_HLEN + 4)
376
377 /*
378 * There seems to be a magic difference in the effect between 995 and 996
379 * but little difference between 900 and 995 ... no idea why.
380 *
381 * There is now a default set of tuning parameters which is set, depending
382 * on whether or not the user enables Jumbo frames. It's assumed that if
383 * Jumbo frames are enabled, the user wants optimal tuning for that case.
384 */
385 #define DEF_TX_COAL 400 /* 996 */
386 #define DEF_TX_MAX_DESC 60 /* was 40 */
387 #define DEF_RX_COAL 120 /* 1000 */
388 #define DEF_RX_MAX_DESC 25
389 #define DEF_TX_RATIO 21 /* 24 */
390
391 #define DEF_JUMBO_TX_COAL 20
392 #define DEF_JUMBO_TX_MAX_DESC 60
393 #define DEF_JUMBO_RX_COAL 30
394 #define DEF_JUMBO_RX_MAX_DESC 6
395 #define DEF_JUMBO_TX_RATIO 21
396
397 #if tigon2FwReleaseLocal < 20001118
398 /*
399 * Standard firmware and early modifications duplicate
400 * IRQ load without this flag (coal timer is never reset).
401 * Note that with this flag tx_coal should be less than
402 * time to xmit full tx ring.
403 * 400usec is not so bad for tx ring size of 128.
404 */
405 #define TX_COAL_INTS_ONLY 1 /* worth it */
406 #else
407 /*
408 * With modified firmware, this is not necessary, but still useful.
409 */
410 #define TX_COAL_INTS_ONLY 1
411 #endif
412
413 #define DEF_TRACE 0
414 #define DEF_STAT (2 * TICKS_PER_SEC)
415
416
417 static int link[ACE_MAX_MOD_PARMS];
418 static int trace[ACE_MAX_MOD_PARMS];
419 static int tx_coal_tick[ACE_MAX_MOD_PARMS];
420 static int rx_coal_tick[ACE_MAX_MOD_PARMS];
421 static int max_tx_desc[ACE_MAX_MOD_PARMS];
422 static int max_rx_desc[ACE_MAX_MOD_PARMS];
423 static int tx_ratio[ACE_MAX_MOD_PARMS];
424 static int dis_pci_mem_inval[ACE_MAX_MOD_PARMS] = {1, 1, 1, 1, 1, 1, 1, 1};
425
426 MODULE_AUTHOR("Jes Sorensen <jes@trained-monkey.org>");
427 MODULE_LICENSE("GPL");
428 MODULE_DESCRIPTION("AceNIC/3C985/GA620 Gigabit Ethernet driver");
429
430 module_param_array(link, int, NULL, 0);
431 module_param_array(trace, int, NULL, 0);
432 module_param_array(tx_coal_tick, int, NULL, 0);
433 module_param_array(max_tx_desc, int, NULL, 0);
434 module_param_array(rx_coal_tick, int, NULL, 0);
435 module_param_array(max_rx_desc, int, NULL, 0);
436 module_param_array(tx_ratio, int, NULL, 0);
437 MODULE_PARM_DESC(link, "AceNIC/3C985/NetGear link state");
438 MODULE_PARM_DESC(trace, "AceNIC/3C985/NetGear firmware trace level");
439 MODULE_PARM_DESC(tx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first tx descriptor arrives");
440 MODULE_PARM_DESC(max_tx_desc, "AceNIC/3C985/GA620 max number of transmit descriptors to wait");
441 MODULE_PARM_DESC(rx_coal_tick, "AceNIC/3C985/GA620 max clock ticks to wait from first rx descriptor arrives");
442 MODULE_PARM_DESC(max_rx_desc, "AceNIC/3C985/GA620 max number of receive descriptors to wait");
443 MODULE_PARM_DESC(tx_ratio, "AceNIC/3C985/GA620 ratio of NIC memory used for TX/RX descriptors (range 0-63)");
444
445
446 static char version[] __devinitdata =
447 "acenic.c: v0.92 08/05/2002 Jes Sorensen, linux-acenic@SunSITE.dk\n"
448 " http://home.cern.ch/~jes/gige/acenic.html\n";
449
450 static int ace_get_settings(struct net_device *, struct ethtool_cmd *);
451 static int ace_set_settings(struct net_device *, struct ethtool_cmd *);
452 static void ace_get_drvinfo(struct net_device *, struct ethtool_drvinfo *);
453
454 static struct ethtool_ops ace_ethtool_ops = {
455 .get_settings = ace_get_settings,
456 .set_settings = ace_set_settings,
457 .get_drvinfo = ace_get_drvinfo,
458 };
459
460 static void ace_watchdog(struct net_device *dev);
461
462 static int __devinit acenic_probe_one(struct pci_dev *pdev,
463 const struct pci_device_id *id)
464 {
465 struct net_device *dev;
466 struct ace_private *ap;
467 static int boards_found;
468
469 dev = alloc_etherdev(sizeof(struct ace_private));
470 if (dev == NULL) {
471 printk(KERN_ERR "acenic: Unable to allocate "
472 "net_device structure!\n");
473 return -ENOMEM;
474 }
475
476 SET_MODULE_OWNER(dev);
477 SET_NETDEV_DEV(dev, &pdev->dev);
478
479 ap = dev->priv;
480 ap->pdev = pdev;
481 ap->name = pci_name(pdev);
482
483 dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
484 #if ACENIC_DO_VLAN
485 dev->features |= NETIF_F_HW_VLAN_TX | NETIF_F_HW_VLAN_RX;
486 dev->vlan_rx_register = ace_vlan_rx_register;
487 dev->vlan_rx_kill_vid = ace_vlan_rx_kill_vid;
488 #endif
489 if (1) {
490 dev->tx_timeout = &ace_watchdog;
491 dev->watchdog_timeo = 5*HZ;
492 }
493
494 dev->open = &ace_open;
495 dev->stop = &ace_close;
496 dev->hard_start_xmit = &ace_start_xmit;
497 dev->get_stats = &ace_get_stats;
498 dev->set_multicast_list = &ace_set_multicast_list;
499 SET_ETHTOOL_OPS(dev, &ace_ethtool_ops);
500 dev->set_mac_address = &ace_set_mac_addr;
501 dev->change_mtu = &ace_change_mtu;
502
503 /* we only display this string ONCE */
504 if (!boards_found)
505 printk(version);
506
507 if (pci_enable_device(pdev))
508 goto fail_free_netdev;
509
510 /*
511 * Enable master mode before we start playing with the
512 * pci_command word since pci_set_master() will modify
513 * it.
514 */
515 pci_set_master(pdev);
516
517 pci_read_config_word(pdev, PCI_COMMAND, &ap->pci_command);
518
519 /* OpenFirmware on Mac's does not set this - DOH.. */
520 if (!(ap->pci_command & PCI_COMMAND_MEMORY)) {
521 printk(KERN_INFO "%s: Enabling PCI Memory Mapped "
522 "access - was not enabled by BIOS/Firmware\n",
523 ap->name);
524 ap->pci_command = ap->pci_command | PCI_COMMAND_MEMORY;
525 pci_write_config_word(ap->pdev, PCI_COMMAND,
526 ap->pci_command);
527 wmb();
528 }
529
530 pci_read_config_byte(pdev, PCI_LATENCY_TIMER, &ap->pci_latency);
531 if (ap->pci_latency <= 0x40) {
532 ap->pci_latency = 0x40;
533 pci_write_config_byte(pdev, PCI_LATENCY_TIMER, ap->pci_latency);
534 }
535
536 /*
537 * Remap the regs into kernel space - this is abuse of
538 * dev->base_addr since it was means for I/O port
539 * addresses but who gives a damn.
540 */
541 dev->base_addr = pci_resource_start(pdev, 0);
542 ap->regs = ioremap(dev->base_addr, 0x4000);
543 if (!ap->regs) {
544 printk(KERN_ERR "%s: Unable to map I/O register, "
545 "AceNIC %i will be disabled.\n",
546 ap->name, boards_found);
547 goto fail_free_netdev;
548 }
549
550 switch(pdev->vendor) {
551 case PCI_VENDOR_ID_ALTEON:
552 if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9100T) {
553 printk(KERN_INFO "%s: Farallon PN9100-T ",
554 ap->name);
555 } else {
556 printk(KERN_INFO "%s: Alteon AceNIC ",
557 ap->name);
558 }
559 break;
560 case PCI_VENDOR_ID_3COM:
561 printk(KERN_INFO "%s: 3Com 3C985 ", ap->name);
562 break;
563 case PCI_VENDOR_ID_NETGEAR:
564 printk(KERN_INFO "%s: NetGear GA620 ", ap->name);
565 break;
566 case PCI_VENDOR_ID_DEC:
567 if (pdev->device == PCI_DEVICE_ID_FARALLON_PN9000SX) {
568 printk(KERN_INFO "%s: Farallon PN9000-SX ",
569 ap->name);
570 break;
571 }
572 case PCI_VENDOR_ID_SGI:
573 printk(KERN_INFO "%s: SGI AceNIC ", ap->name);
574 break;
575 default:
576 printk(KERN_INFO "%s: Unknown AceNIC ", ap->name);
577 break;
578 }
579
580 printk("Gigabit Ethernet at 0x%08lx, ", dev->base_addr);
581 printk("irq %d\n", pdev->irq);
582
583 #ifdef CONFIG_ACENIC_OMIT_TIGON_I
584 if ((readl(&ap->regs->HostCtrl) >> 28) == 4) {
585 printk(KERN_ERR "%s: Driver compiled without Tigon I"
586 " support - NIC disabled\n", dev->name);
587 goto fail_uninit;
588 }
589 #endif
590
591 if (ace_allocate_descriptors(dev))
592 goto fail_free_netdev;
593
594 #ifdef MODULE
595 if (boards_found >= ACE_MAX_MOD_PARMS)
596 ap->board_idx = BOARD_IDX_OVERFLOW;
597 else
598 ap->board_idx = boards_found;
599 #else
600 ap->board_idx = BOARD_IDX_STATIC;
601 #endif
602
603 if (ace_init(dev))
604 goto fail_free_netdev;
605
606 if (register_netdev(dev)) {
607 printk(KERN_ERR "acenic: device registration failed\n");
608 goto fail_uninit;
609 }
610 ap->name = dev->name;
611
612 if (ap->pci_using_dac)
613 dev->features |= NETIF_F_HIGHDMA;
614
615 pci_set_drvdata(pdev, dev);
616
617 boards_found++;
618 return 0;
619
620 fail_uninit:
621 ace_init_cleanup(dev);
622 fail_free_netdev:
623 free_netdev(dev);
624 return -ENODEV;
625 }
626
627 static void __devexit acenic_remove_one(struct pci_dev *pdev)
628 {
629 struct net_device *dev = pci_get_drvdata(pdev);
630 struct ace_private *ap = netdev_priv(dev);
631 struct ace_regs __iomem *regs = ap->regs;
632 short i;
633
634 unregister_netdev(dev);
635
636 writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
637 if (ap->version >= 2)
638 writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
639
640 /*
641 * This clears any pending interrupts
642 */
643 writel(1, &regs->Mb0Lo);
644 readl(&regs->CpuCtrl); /* flush */
645
646 /*
647 * Make sure no other CPUs are processing interrupts
648 * on the card before the buffers are being released.
649 * Otherwise one might experience some `interesting'
650 * effects.
651 *
652 * Then release the RX buffers - jumbo buffers were
653 * already released in ace_close().
654 */
655 ace_sync_irq(dev->irq);
656
657 for (i = 0; i < RX_STD_RING_ENTRIES; i++) {
658 struct sk_buff *skb = ap->skb->rx_std_skbuff[i].skb;
659
660 if (skb) {
661 struct ring_info *ringp;
662 dma_addr_t mapping;
663
664 ringp = &ap->skb->rx_std_skbuff[i];
665 mapping = pci_unmap_addr(ringp, mapping);
666 pci_unmap_page(ap->pdev, mapping,
667 ACE_STD_BUFSIZE,
668 PCI_DMA_FROMDEVICE);
669
670 ap->rx_std_ring[i].size = 0;
671 ap->skb->rx_std_skbuff[i].skb = NULL;
672 dev_kfree_skb(skb);
673 }
674 }
675
676 if (ap->version >= 2) {
677 for (i = 0; i < RX_MINI_RING_ENTRIES; i++) {
678 struct sk_buff *skb = ap->skb->rx_mini_skbuff[i].skb;
679
680 if (skb) {
681 struct ring_info *ringp;
682 dma_addr_t mapping;
683
684 ringp = &ap->skb->rx_mini_skbuff[i];
685 mapping = pci_unmap_addr(ringp,mapping);
686 pci_unmap_page(ap->pdev, mapping,
687 ACE_MINI_BUFSIZE,
688 PCI_DMA_FROMDEVICE);
689
690 ap->rx_mini_ring[i].size = 0;
691 ap->skb->rx_mini_skbuff[i].skb = NULL;
692 dev_kfree_skb(skb);
693 }
694 }
695 }
696
697 for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
698 struct sk_buff *skb = ap->skb->rx_jumbo_skbuff[i].skb;
699 if (skb) {
700 struct ring_info *ringp;
701 dma_addr_t mapping;
702
703 ringp = &ap->skb->rx_jumbo_skbuff[i];
704 mapping = pci_unmap_addr(ringp, mapping);
705 pci_unmap_page(ap->pdev, mapping,
706 ACE_JUMBO_BUFSIZE,
707 PCI_DMA_FROMDEVICE);
708
709 ap->rx_jumbo_ring[i].size = 0;
710 ap->skb->rx_jumbo_skbuff[i].skb = NULL;
711 dev_kfree_skb(skb);
712 }
713 }
714
715 ace_init_cleanup(dev);
716 free_netdev(dev);
717 }
718
719 static struct pci_driver acenic_pci_driver = {
720 .name = "acenic",
721 .id_table = acenic_pci_tbl,
722 .probe = acenic_probe_one,
723 .remove = __devexit_p(acenic_remove_one),
724 };
725
726 static int __init acenic_init(void)
727 {
728 return pci_module_init(&acenic_pci_driver);
729 }
730
731 static void __exit acenic_exit(void)
732 {
733 pci_unregister_driver(&acenic_pci_driver);
734 }
735
736 module_init(acenic_init);
737 module_exit(acenic_exit);
738
739 static void ace_free_descriptors(struct net_device *dev)
740 {
741 struct ace_private *ap = netdev_priv(dev);
742 int size;
743
744 if (ap->rx_std_ring != NULL) {
745 size = (sizeof(struct rx_desc) *
746 (RX_STD_RING_ENTRIES +
747 RX_JUMBO_RING_ENTRIES +
748 RX_MINI_RING_ENTRIES +
749 RX_RETURN_RING_ENTRIES));
750 pci_free_consistent(ap->pdev, size, ap->rx_std_ring,
751 ap->rx_ring_base_dma);
752 ap->rx_std_ring = NULL;
753 ap->rx_jumbo_ring = NULL;
754 ap->rx_mini_ring = NULL;
755 ap->rx_return_ring = NULL;
756 }
757 if (ap->evt_ring != NULL) {
758 size = (sizeof(struct event) * EVT_RING_ENTRIES);
759 pci_free_consistent(ap->pdev, size, ap->evt_ring,
760 ap->evt_ring_dma);
761 ap->evt_ring = NULL;
762 }
763 if (ap->tx_ring != NULL && !ACE_IS_TIGON_I(ap)) {
764 size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
765 pci_free_consistent(ap->pdev, size, ap->tx_ring,
766 ap->tx_ring_dma);
767 }
768 ap->tx_ring = NULL;
769
770 if (ap->evt_prd != NULL) {
771 pci_free_consistent(ap->pdev, sizeof(u32),
772 (void *)ap->evt_prd, ap->evt_prd_dma);
773 ap->evt_prd = NULL;
774 }
775 if (ap->rx_ret_prd != NULL) {
776 pci_free_consistent(ap->pdev, sizeof(u32),
777 (void *)ap->rx_ret_prd,
778 ap->rx_ret_prd_dma);
779 ap->rx_ret_prd = NULL;
780 }
781 if (ap->tx_csm != NULL) {
782 pci_free_consistent(ap->pdev, sizeof(u32),
783 (void *)ap->tx_csm, ap->tx_csm_dma);
784 ap->tx_csm = NULL;
785 }
786 }
787
788
789 static int ace_allocate_descriptors(struct net_device *dev)
790 {
791 struct ace_private *ap = netdev_priv(dev);
792 int size;
793
794 size = (sizeof(struct rx_desc) *
795 (RX_STD_RING_ENTRIES +
796 RX_JUMBO_RING_ENTRIES +
797 RX_MINI_RING_ENTRIES +
798 RX_RETURN_RING_ENTRIES));
799
800 ap->rx_std_ring = pci_alloc_consistent(ap->pdev, size,
801 &ap->rx_ring_base_dma);
802 if (ap->rx_std_ring == NULL)
803 goto fail;
804
805 ap->rx_jumbo_ring = ap->rx_std_ring + RX_STD_RING_ENTRIES;
806 ap->rx_mini_ring = ap->rx_jumbo_ring + RX_JUMBO_RING_ENTRIES;
807 ap->rx_return_ring = ap->rx_mini_ring + RX_MINI_RING_ENTRIES;
808
809 size = (sizeof(struct event) * EVT_RING_ENTRIES);
810
811 ap->evt_ring = pci_alloc_consistent(ap->pdev, size, &ap->evt_ring_dma);
812
813 if (ap->evt_ring == NULL)
814 goto fail;
815
816 /*
817 * Only allocate a host TX ring for the Tigon II, the Tigon I
818 * has to use PCI registers for this ;-(
819 */
820 if (!ACE_IS_TIGON_I(ap)) {
821 size = (sizeof(struct tx_desc) * MAX_TX_RING_ENTRIES);
822
823 ap->tx_ring = pci_alloc_consistent(ap->pdev, size,
824 &ap->tx_ring_dma);
825
826 if (ap->tx_ring == NULL)
827 goto fail;
828 }
829
830 ap->evt_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
831 &ap->evt_prd_dma);
832 if (ap->evt_prd == NULL)
833 goto fail;
834
835 ap->rx_ret_prd = pci_alloc_consistent(ap->pdev, sizeof(u32),
836 &ap->rx_ret_prd_dma);
837 if (ap->rx_ret_prd == NULL)
838 goto fail;
839
840 ap->tx_csm = pci_alloc_consistent(ap->pdev, sizeof(u32),
841 &ap->tx_csm_dma);
842 if (ap->tx_csm == NULL)
843 goto fail;
844
845 return 0;
846
847 fail:
848 /* Clean up. */
849 ace_init_cleanup(dev);
850 return 1;
851 }
852
853
854 /*
855 * Generic cleanup handling data allocated during init. Used when the
856 * module is unloaded or if an error occurs during initialization
857 */
858 static void ace_init_cleanup(struct net_device *dev)
859 {
860 struct ace_private *ap;
861
862 ap = netdev_priv(dev);
863
864 ace_free_descriptors(dev);
865
866 if (ap->info)
867 pci_free_consistent(ap->pdev, sizeof(struct ace_info),
868 ap->info, ap->info_dma);
869 kfree(ap->skb);
870 kfree(ap->trace_buf);
871
872 if (dev->irq)
873 free_irq(dev->irq, dev);
874
875 iounmap(ap->regs);
876 }
877
878
879 /*
880 * Commands are considered to be slow.
881 */
882 static inline void ace_issue_cmd(struct ace_regs __iomem *regs, struct cmd *cmd)
883 {
884 u32 idx;
885
886 idx = readl(&regs->CmdPrd);
887
888 writel(*(u32 *)(cmd), &regs->CmdRng[idx]);
889 idx = (idx + 1) % CMD_RING_ENTRIES;
890
891 writel(idx, &regs->CmdPrd);
892 }
893
894
895 static int __devinit ace_init(struct net_device *dev)
896 {
897 struct ace_private *ap;
898 struct ace_regs __iomem *regs;
899 struct ace_info *info = NULL;
900 struct pci_dev *pdev;
901 unsigned long myjif;
902 u64 tmp_ptr;
903 u32 tig_ver, mac1, mac2, tmp, pci_state;
904 int board_idx, ecode = 0;
905 short i;
906 unsigned char cache_size;
907
908 ap = netdev_priv(dev);
909 regs = ap->regs;
910
911 board_idx = ap->board_idx;
912
913 /*
914 * aman@sgi.com - its useful to do a NIC reset here to
915 * address the `Firmware not running' problem subsequent
916 * to any crashes involving the NIC
917 */
918 writel(HW_RESET | (HW_RESET << 24), &regs->HostCtrl);
919 readl(&regs->HostCtrl); /* PCI write posting */
920 udelay(5);
921
922 /*
923 * Don't access any other registers before this point!
924 */
925 #ifdef __BIG_ENDIAN
926 /*
927 * This will most likely need BYTE_SWAP once we switch
928 * to using __raw_writel()
929 */
930 writel((WORD_SWAP | CLR_INT | ((WORD_SWAP | CLR_INT) << 24)),
931 &regs->HostCtrl);
932 #else
933 writel((CLR_INT | WORD_SWAP | ((CLR_INT | WORD_SWAP) << 24)),
934 &regs->HostCtrl);
935 #endif
936 readl(&regs->HostCtrl); /* PCI write posting */
937
938 /*
939 * Stop the NIC CPU and clear pending interrupts
940 */
941 writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
942 readl(&regs->CpuCtrl); /* PCI write posting */
943 writel(0, &regs->Mb0Lo);
944
945 tig_ver = readl(&regs->HostCtrl) >> 28;
946
947 switch(tig_ver){
948 #ifndef CONFIG_ACENIC_OMIT_TIGON_I
949 case 4:
950 case 5:
951 printk(KERN_INFO " Tigon I (Rev. %i), Firmware: %i.%i.%i, ",
952 tig_ver, tigonFwReleaseMajor, tigonFwReleaseMinor,
953 tigonFwReleaseFix);
954 writel(0, &regs->LocalCtrl);
955 ap->version = 1;
956 ap->tx_ring_entries = TIGON_I_TX_RING_ENTRIES;
957 break;
958 #endif
959 case 6:
960 printk(KERN_INFO " Tigon II (Rev. %i), Firmware: %i.%i.%i, ",
961 tig_ver, tigon2FwReleaseMajor, tigon2FwReleaseMinor,
962 tigon2FwReleaseFix);
963 writel(readl(&regs->CpuBCtrl) | CPU_HALT, &regs->CpuBCtrl);
964 readl(&regs->CpuBCtrl); /* PCI write posting */
965 /*
966 * The SRAM bank size does _not_ indicate the amount
967 * of memory on the card, it controls the _bank_ size!
968 * Ie. a 1MB AceNIC will have two banks of 512KB.
969 */
970 writel(SRAM_BANK_512K, &regs->LocalCtrl);
971 writel(SYNC_SRAM_TIMING, &regs->MiscCfg);
972 ap->version = 2;
973 ap->tx_ring_entries = MAX_TX_RING_ENTRIES;
974 break;
975 default:
976 printk(KERN_WARNING " Unsupported Tigon version detected "
977 "(%i)\n", tig_ver);
978 ecode = -ENODEV;
979 goto init_error;
980 }
981
982 /*
983 * ModeStat _must_ be set after the SRAM settings as this change
984 * seems to corrupt the ModeStat and possible other registers.
985 * The SRAM settings survive resets and setting it to the same
986 * value a second time works as well. This is what caused the
987 * `Firmware not running' problem on the Tigon II.
988 */
989 #ifdef __BIG_ENDIAN
990 writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL | ACE_BYTE_SWAP_BD |
991 ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
992 #else
993 writel(ACE_BYTE_SWAP_DMA | ACE_WARN | ACE_FATAL |
994 ACE_WORD_SWAP_BD | ACE_NO_JUMBO_FRAG, &regs->ModeStat);
995 #endif
996 readl(&regs->ModeStat); /* PCI write posting */
997
998 mac1 = 0;
999 for(i = 0; i < 4; i++) {
1000 int tmp;
1001
1002 mac1 = mac1 << 8;
1003 tmp = read_eeprom_byte(dev, 0x8c+i);
1004 if (tmp < 0) {
1005 ecode = -EIO;
1006 goto init_error;
1007 } else
1008 mac1 |= (tmp & 0xff);
1009 }
1010 mac2 = 0;
1011 for(i = 4; i < 8; i++) {
1012 int tmp;
1013
1014 mac2 = mac2 << 8;
1015 tmp = read_eeprom_byte(dev, 0x8c+i);
1016 if (tmp < 0) {
1017 ecode = -EIO;
1018 goto init_error;
1019 } else
1020 mac2 |= (tmp & 0xff);
1021 }
1022
1023 writel(mac1, &regs->MacAddrHi);
1024 writel(mac2, &regs->MacAddrLo);
1025
1026 printk("MAC: %02x:%02x:%02x:%02x:%02x:%02x\n",
1027 (mac1 >> 8) & 0xff, mac1 & 0xff, (mac2 >> 24) &0xff,
1028 (mac2 >> 16) & 0xff, (mac2 >> 8) & 0xff, mac2 & 0xff);
1029
1030 dev->dev_addr[0] = (mac1 >> 8) & 0xff;
1031 dev->dev_addr[1] = mac1 & 0xff;
1032 dev->dev_addr[2] = (mac2 >> 24) & 0xff;
1033 dev->dev_addr[3] = (mac2 >> 16) & 0xff;
1034 dev->dev_addr[4] = (mac2 >> 8) & 0xff;
1035 dev->dev_addr[5] = mac2 & 0xff;
1036
1037 /*
1038 * Looks like this is necessary to deal with on all architectures,
1039 * even this %$#%$# N440BX Intel based thing doesn't get it right.
1040 * Ie. having two NICs in the machine, one will have the cache
1041 * line set at boot time, the other will not.
1042 */
1043 pdev = ap->pdev;
1044 pci_read_config_byte(pdev, PCI_CACHE_LINE_SIZE, &cache_size);
1045 cache_size <<= 2;
1046 if (cache_size != SMP_CACHE_BYTES) {
1047 printk(KERN_INFO " PCI cache line size set incorrectly "
1048 "(%i bytes) by BIOS/FW, ", cache_size);
1049 if (cache_size > SMP_CACHE_BYTES)
1050 printk("expecting %i\n", SMP_CACHE_BYTES);
1051 else {
1052 printk("correcting to %i\n", SMP_CACHE_BYTES);
1053 pci_write_config_byte(pdev, PCI_CACHE_LINE_SIZE,
1054 SMP_CACHE_BYTES >> 2);
1055 }
1056 }
1057
1058 pci_state = readl(&regs->PciState);
1059 printk(KERN_INFO " PCI bus width: %i bits, speed: %iMHz, "
1060 "latency: %i clks\n",
1061 (pci_state & PCI_32BIT) ? 32 : 64,
1062 (pci_state & PCI_66MHZ) ? 66 : 33,
1063 ap->pci_latency);
1064
1065 /*
1066 * Set the max DMA transfer size. Seems that for most systems
1067 * the performance is better when no MAX parameter is
1068 * set. However for systems enabling PCI write and invalidate,
1069 * DMA writes must be set to the L1 cache line size to get
1070 * optimal performance.
1071 *
1072 * The default is now to turn the PCI write and invalidate off
1073 * - that is what Alteon does for NT.
1074 */
1075 tmp = READ_CMD_MEM | WRITE_CMD_MEM;
1076 if (ap->version >= 2) {
1077 tmp |= (MEM_READ_MULTIPLE | (pci_state & PCI_66MHZ));
1078 /*
1079 * Tuning parameters only supported for 8 cards
1080 */
1081 if (board_idx == BOARD_IDX_OVERFLOW ||
1082 dis_pci_mem_inval[board_idx]) {
1083 if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1084 ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1085 pci_write_config_word(pdev, PCI_COMMAND,
1086 ap->pci_command);
1087 printk(KERN_INFO " Disabling PCI memory "
1088 "write and invalidate\n");
1089 }
1090 } else if (ap->pci_command & PCI_COMMAND_INVALIDATE) {
1091 printk(KERN_INFO " PCI memory write & invalidate "
1092 "enabled by BIOS, enabling counter measures\n");
1093
1094 switch(SMP_CACHE_BYTES) {
1095 case 16:
1096 tmp |= DMA_WRITE_MAX_16;
1097 break;
1098 case 32:
1099 tmp |= DMA_WRITE_MAX_32;
1100 break;
1101 case 64:
1102 tmp |= DMA_WRITE_MAX_64;
1103 break;
1104 case 128:
1105 tmp |= DMA_WRITE_MAX_128;
1106 break;
1107 default:
1108 printk(KERN_INFO " Cache line size %i not "
1109 "supported, PCI write and invalidate "
1110 "disabled\n", SMP_CACHE_BYTES);
1111 ap->pci_command &= ~PCI_COMMAND_INVALIDATE;
1112 pci_write_config_word(pdev, PCI_COMMAND,
1113 ap->pci_command);
1114 }
1115 }
1116 }
1117
1118 #ifdef __sparc__
1119 /*
1120 * On this platform, we know what the best dma settings
1121 * are. We use 64-byte maximum bursts, because if we
1122 * burst larger than the cache line size (or even cross
1123 * a 64byte boundary in a single burst) the UltraSparc
1124 * PCI controller will disconnect at 64-byte multiples.
1125 *
1126 * Read-multiple will be properly enabled above, and when
1127 * set will give the PCI controller proper hints about
1128 * prefetching.
1129 */
1130 tmp &= ~DMA_READ_WRITE_MASK;
1131 tmp |= DMA_READ_MAX_64;
1132 tmp |= DMA_WRITE_MAX_64;
1133 #endif
1134 #ifdef __alpha__
1135 tmp &= ~DMA_READ_WRITE_MASK;
1136 tmp |= DMA_READ_MAX_128;
1137 /*
1138 * All the docs say MUST NOT. Well, I did.
1139 * Nothing terrible happens, if we load wrong size.
1140 * Bit w&i still works better!
1141 */
1142 tmp |= DMA_WRITE_MAX_128;
1143 #endif
1144 writel(tmp, &regs->PciState);
1145
1146 #if 0
1147 /*
1148 * The Host PCI bus controller driver has to set FBB.
1149 * If all devices on that PCI bus support FBB, then the controller
1150 * can enable FBB support in the Host PCI Bus controller (or on
1151 * the PCI-PCI bridge if that applies).
1152 * -ggg
1153 */
1154 /*
1155 * I have received reports from people having problems when this
1156 * bit is enabled.
1157 */
1158 if (!(ap->pci_command & PCI_COMMAND_FAST_BACK)) {
1159 printk(KERN_INFO " Enabling PCI Fast Back to Back\n");
1160 ap->pci_command |= PCI_COMMAND_FAST_BACK;
1161 pci_write_config_word(pdev, PCI_COMMAND, ap->pci_command);
1162 }
1163 #endif
1164
1165 /*
1166 * Configure DMA attributes.
1167 */
1168 if (!pci_set_dma_mask(pdev, DMA_64BIT_MASK)) {
1169 ap->pci_using_dac = 1;
1170 } else if (!pci_set_dma_mask(pdev, DMA_32BIT_MASK)) {
1171 ap->pci_using_dac = 0;
1172 } else {
1173 ecode = -ENODEV;
1174 goto init_error;
1175 }
1176
1177 /*
1178 * Initialize the generic info block and the command+event rings
1179 * and the control blocks for the transmit and receive rings
1180 * as they need to be setup once and for all.
1181 */
1182 if (!(info = pci_alloc_consistent(ap->pdev, sizeof(struct ace_info),
1183 &ap->info_dma))) {
1184 ecode = -EAGAIN;
1185 goto init_error;
1186 }
1187 ap->info = info;
1188
1189 /*
1190 * Get the memory for the skb rings.
1191 */
1192 if (!(ap->skb = kmalloc(sizeof(struct ace_skb), GFP_KERNEL))) {
1193 ecode = -EAGAIN;
1194 goto init_error;
1195 }
1196
1197 ecode = request_irq(pdev->irq, ace_interrupt, IRQF_SHARED,
1198 DRV_NAME, dev);
1199 if (ecode) {
1200 printk(KERN_WARNING "%s: Requested IRQ %d is busy\n",
1201 DRV_NAME, pdev->irq);
1202 goto init_error;
1203 } else
1204 dev->irq = pdev->irq;
1205
1206 #ifdef INDEX_DEBUG
1207 spin_lock_init(&ap->debug_lock);
1208 ap->last_tx = ACE_TX_RING_ENTRIES(ap) - 1;
1209 ap->last_std_rx = 0;
1210 ap->last_mini_rx = 0;
1211 #endif
1212
1213 memset(ap->info, 0, sizeof(struct ace_info));
1214 memset(ap->skb, 0, sizeof(struct ace_skb));
1215
1216 ace_load_firmware(dev);
1217 ap->fw_running = 0;
1218
1219 tmp_ptr = ap->info_dma;
1220 writel(tmp_ptr >> 32, &regs->InfoPtrHi);
1221 writel(tmp_ptr & 0xffffffff, &regs->InfoPtrLo);
1222
1223 memset(ap->evt_ring, 0, EVT_RING_ENTRIES * sizeof(struct event));
1224
1225 set_aceaddr(&info->evt_ctrl.rngptr, ap->evt_ring_dma);
1226 info->evt_ctrl.flags = 0;
1227
1228 *(ap->evt_prd) = 0;
1229 wmb();
1230 set_aceaddr(&info->evt_prd_ptr, ap->evt_prd_dma);
1231 writel(0, &regs->EvtCsm);
1232
1233 set_aceaddr(&info->cmd_ctrl.rngptr, 0x100);
1234 info->cmd_ctrl.flags = 0;
1235 info->cmd_ctrl.max_len = 0;
1236
1237 for (i = 0; i < CMD_RING_ENTRIES; i++)
1238 writel(0, &regs->CmdRng[i]);
1239
1240 writel(0, &regs->CmdPrd);
1241 writel(0, &regs->CmdCsm);
1242
1243 tmp_ptr = ap->info_dma;
1244 tmp_ptr += (unsigned long) &(((struct ace_info *)0)->s.stats);
1245 set_aceaddr(&info->stats2_ptr, (dma_addr_t) tmp_ptr);
1246
1247 set_aceaddr(&info->rx_std_ctrl.rngptr, ap->rx_ring_base_dma);
1248 info->rx_std_ctrl.max_len = ACE_STD_BUFSIZE;
1249 info->rx_std_ctrl.flags =
1250 RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | ACE_RCB_VLAN_FLAG;
1251
1252 memset(ap->rx_std_ring, 0,
1253 RX_STD_RING_ENTRIES * sizeof(struct rx_desc));
1254
1255 for (i = 0; i < RX_STD_RING_ENTRIES; i++)
1256 ap->rx_std_ring[i].flags = BD_FLG_TCP_UDP_SUM;
1257
1258 ap->rx_std_skbprd = 0;
1259 atomic_set(&ap->cur_rx_bufs, 0);
1260
1261 set_aceaddr(&info->rx_jumbo_ctrl.rngptr,
1262 (ap->rx_ring_base_dma +
1263 (sizeof(struct rx_desc) * RX_STD_RING_ENTRIES)));
1264 info->rx_jumbo_ctrl.max_len = 0;
1265 info->rx_jumbo_ctrl.flags =
1266 RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | ACE_RCB_VLAN_FLAG;
1267
1268 memset(ap->rx_jumbo_ring, 0,
1269 RX_JUMBO_RING_ENTRIES * sizeof(struct rx_desc));
1270
1271 for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++)
1272 ap->rx_jumbo_ring[i].flags = BD_FLG_TCP_UDP_SUM | BD_FLG_JUMBO;
1273
1274 ap->rx_jumbo_skbprd = 0;
1275 atomic_set(&ap->cur_jumbo_bufs, 0);
1276
1277 memset(ap->rx_mini_ring, 0,
1278 RX_MINI_RING_ENTRIES * sizeof(struct rx_desc));
1279
1280 if (ap->version >= 2) {
1281 set_aceaddr(&info->rx_mini_ctrl.rngptr,
1282 (ap->rx_ring_base_dma +
1283 (sizeof(struct rx_desc) *
1284 (RX_STD_RING_ENTRIES +
1285 RX_JUMBO_RING_ENTRIES))));
1286 info->rx_mini_ctrl.max_len = ACE_MINI_SIZE;
1287 info->rx_mini_ctrl.flags =
1288 RCB_FLG_TCP_UDP_SUM|RCB_FLG_NO_PSEUDO_HDR|ACE_RCB_VLAN_FLAG;
1289
1290 for (i = 0; i < RX_MINI_RING_ENTRIES; i++)
1291 ap->rx_mini_ring[i].flags =
1292 BD_FLG_TCP_UDP_SUM | BD_FLG_MINI;
1293 } else {
1294 set_aceaddr(&info->rx_mini_ctrl.rngptr, 0);
1295 info->rx_mini_ctrl.flags = RCB_FLG_RNG_DISABLE;
1296 info->rx_mini_ctrl.max_len = 0;
1297 }
1298
1299 ap->rx_mini_skbprd = 0;
1300 atomic_set(&ap->cur_mini_bufs, 0);
1301
1302 set_aceaddr(&info->rx_return_ctrl.rngptr,
1303 (ap->rx_ring_base_dma +
1304 (sizeof(struct rx_desc) *
1305 (RX_STD_RING_ENTRIES +
1306 RX_JUMBO_RING_ENTRIES +
1307 RX_MINI_RING_ENTRIES))));
1308 info->rx_return_ctrl.flags = 0;
1309 info->rx_return_ctrl.max_len = RX_RETURN_RING_ENTRIES;
1310
1311 memset(ap->rx_return_ring, 0,
1312 RX_RETURN_RING_ENTRIES * sizeof(struct rx_desc));
1313
1314 set_aceaddr(&info->rx_ret_prd_ptr, ap->rx_ret_prd_dma);
1315 *(ap->rx_ret_prd) = 0;
1316
1317 writel(TX_RING_BASE, &regs->WinBase);
1318
1319 if (ACE_IS_TIGON_I(ap)) {
1320 ap->tx_ring = (struct tx_desc *) regs->Window;
1321 for (i = 0; i < (TIGON_I_TX_RING_ENTRIES
1322 * sizeof(struct tx_desc)) / sizeof(u32); i++)
1323 writel(0, (void __iomem *)ap->tx_ring + i * 4);
1324
1325 set_aceaddr(&info->tx_ctrl.rngptr, TX_RING_BASE);
1326 } else {
1327 memset(ap->tx_ring, 0,
1328 MAX_TX_RING_ENTRIES * sizeof(struct tx_desc));
1329
1330 set_aceaddr(&info->tx_ctrl.rngptr, ap->tx_ring_dma);
1331 }
1332
1333 info->tx_ctrl.max_len = ACE_TX_RING_ENTRIES(ap);
1334 tmp = RCB_FLG_TCP_UDP_SUM | RCB_FLG_NO_PSEUDO_HDR | ACE_RCB_VLAN_FLAG;
1335
1336 /*
1337 * The Tigon I does not like having the TX ring in host memory ;-(
1338 */
1339 if (!ACE_IS_TIGON_I(ap))
1340 tmp |= RCB_FLG_TX_HOST_RING;
1341 #if TX_COAL_INTS_ONLY
1342 tmp |= RCB_FLG_COAL_INT_ONLY;
1343 #endif
1344 info->tx_ctrl.flags = tmp;
1345
1346 set_aceaddr(&info->tx_csm_ptr, ap->tx_csm_dma);
1347
1348 /*
1349 * Potential item for tuning parameter
1350 */
1351 #if 0 /* NO */
1352 writel(DMA_THRESH_16W, &regs->DmaReadCfg);
1353 writel(DMA_THRESH_16W, &regs->DmaWriteCfg);
1354 #else
1355 writel(DMA_THRESH_8W, &regs->DmaReadCfg);
1356 writel(DMA_THRESH_8W, &regs->DmaWriteCfg);
1357 #endif
1358
1359 writel(0, &regs->MaskInt);
1360 writel(1, &regs->IfIdx);
1361 #if 0
1362 /*
1363 * McKinley boxes do not like us fiddling with AssistState
1364 * this early
1365 */
1366 writel(1, &regs->AssistState);
1367 #endif
1368
1369 writel(DEF_STAT, &regs->TuneStatTicks);
1370 writel(DEF_TRACE, &regs->TuneTrace);
1371
1372 ace_set_rxtx_parms(dev, 0);
1373
1374 if (board_idx == BOARD_IDX_OVERFLOW) {
1375 printk(KERN_WARNING "%s: more than %i NICs detected, "
1376 "ignoring module parameters!\n",
1377 ap->name, ACE_MAX_MOD_PARMS);
1378 } else if (board_idx >= 0) {
1379 if (tx_coal_tick[board_idx])
1380 writel(tx_coal_tick[board_idx],
1381 &regs->TuneTxCoalTicks);
1382 if (max_tx_desc[board_idx])
1383 writel(max_tx_desc[board_idx], &regs->TuneMaxTxDesc);
1384
1385 if (rx_coal_tick[board_idx])
1386 writel(rx_coal_tick[board_idx],
1387 &regs->TuneRxCoalTicks);
1388 if (max_rx_desc[board_idx])
1389 writel(max_rx_desc[board_idx], &regs->TuneMaxRxDesc);
1390
1391 if (trace[board_idx])
1392 writel(trace[board_idx], &regs->TuneTrace);
1393
1394 if ((tx_ratio[board_idx] > 0) && (tx_ratio[board_idx] < 64))
1395 writel(tx_ratio[board_idx], &regs->TxBufRat);
1396 }
1397
1398 /*
1399 * Default link parameters
1400 */
1401 tmp = LNK_ENABLE | LNK_FULL_DUPLEX | LNK_1000MB | LNK_100MB |
1402 LNK_10MB | LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL | LNK_NEGOTIATE;
1403 if(ap->version >= 2)
1404 tmp |= LNK_TX_FLOW_CTL_Y;
1405
1406 /*
1407 * Override link default parameters
1408 */
1409 if ((board_idx >= 0) && link[board_idx]) {
1410 int option = link[board_idx];
1411
1412 tmp = LNK_ENABLE;
1413
1414 if (option & 0x01) {
1415 printk(KERN_INFO "%s: Setting half duplex link\n",
1416 ap->name);
1417 tmp &= ~LNK_FULL_DUPLEX;
1418 }
1419 if (option & 0x02)
1420 tmp &= ~LNK_NEGOTIATE;
1421 if (option & 0x10)
1422 tmp |= LNK_10MB;
1423 if (option & 0x20)
1424 tmp |= LNK_100MB;
1425 if (option & 0x40)
1426 tmp |= LNK_1000MB;
1427 if ((option & 0x70) == 0) {
1428 printk(KERN_WARNING "%s: No media speed specified, "
1429 "forcing auto negotiation\n", ap->name);
1430 tmp |= LNK_NEGOTIATE | LNK_1000MB |
1431 LNK_100MB | LNK_10MB;
1432 }
1433 if ((option & 0x100) == 0)
1434 tmp |= LNK_NEG_FCTL;
1435 else
1436 printk(KERN_INFO "%s: Disabling flow control "
1437 "negotiation\n", ap->name);
1438 if (option & 0x200)
1439 tmp |= LNK_RX_FLOW_CTL_Y;
1440 if ((option & 0x400) && (ap->version >= 2)) {
1441 printk(KERN_INFO "%s: Enabling TX flow control\n",
1442 ap->name);
1443 tmp |= LNK_TX_FLOW_CTL_Y;
1444 }
1445 }
1446
1447 ap->link = tmp;
1448 writel(tmp, &regs->TuneLink);
1449 if (ap->version >= 2)
1450 writel(tmp, &regs->TuneFastLink);
1451
1452 if (ACE_IS_TIGON_I(ap))
1453 writel(tigonFwStartAddr, &regs->Pc);
1454 if (ap->version == 2)
1455 writel(tigon2FwStartAddr, &regs->Pc);
1456
1457 writel(0, &regs->Mb0Lo);
1458
1459 /*
1460 * Set tx_csm before we start receiving interrupts, otherwise
1461 * the interrupt handler might think it is supposed to process
1462 * tx ints before we are up and running, which may cause a null
1463 * pointer access in the int handler.
1464 */
1465 ap->cur_rx = 0;
1466 ap->tx_prd = *(ap->tx_csm) = ap->tx_ret_csm = 0;
1467
1468 wmb();
1469 ace_set_txprd(regs, ap, 0);
1470 writel(0, &regs->RxRetCsm);
1471
1472 /*
1473 * Zero the stats before starting the interface
1474 */
1475 memset(&ap->stats, 0, sizeof(ap->stats));
1476
1477 /*
1478 * Enable DMA engine now.
1479 * If we do this sooner, Mckinley box pukes.
1480 * I assume it's because Tigon II DMA engine wants to check
1481 * *something* even before the CPU is started.
1482 */
1483 writel(1, &regs->AssistState); /* enable DMA */
1484
1485 /*
1486 * Start the NIC CPU
1487 */
1488 writel(readl(&regs->CpuCtrl) & ~(CPU_HALT|CPU_TRACE), &regs->CpuCtrl);
1489 readl(&regs->CpuCtrl);
1490
1491 /*
1492 * Wait for the firmware to spin up - max 3 seconds.
1493 */
1494 myjif = jiffies + 3 * HZ;
1495 while (time_before(jiffies, myjif) && !ap->fw_running)
1496 cpu_relax();
1497
1498 if (!ap->fw_running) {
1499 printk(KERN_ERR "%s: Firmware NOT running!\n", ap->name);
1500
1501 ace_dump_trace(ap);
1502 writel(readl(&regs->CpuCtrl) | CPU_HALT, &regs->CpuCtrl);
1503 readl(&regs->CpuCtrl);
1504
1505 /* aman@sgi.com - account for badly behaving firmware/NIC:
1506 * - have observed that the NIC may continue to generate
1507 * interrupts for some reason; attempt to stop it - halt
1508 * second CPU for Tigon II cards, and also clear Mb0
1509 * - if we're a module, we'll fail to load if this was
1510 * the only GbE card in the system => if the kernel does
1511 * see an interrupt from the NIC, code to handle it is
1512 * gone and OOps! - so free_irq also
1513 */
1514 if (ap->version >= 2)
1515 writel(readl(&regs->CpuBCtrl) | CPU_HALT,
1516 &regs->CpuBCtrl);
1517 writel(0, &regs->Mb0Lo);
1518 readl(&regs->Mb0Lo);
1519
1520 ecode = -EBUSY;
1521 goto init_error;
1522 }
1523
1524 /*
1525 * We load the ring here as there seem to be no way to tell the
1526 * firmware to wipe the ring without re-initializing it.
1527 */
1528 if (!test_and_set_bit(0, &ap->std_refill_busy))
1529 ace_load_std_rx_ring(ap, RX_RING_SIZE);
1530 else
1531 printk(KERN_ERR "%s: Someone is busy refilling the RX ring\n",
1532 ap->name);
1533 if (ap->version >= 2) {
1534 if (!test_and_set_bit(0, &ap->mini_refill_busy))
1535 ace_load_mini_rx_ring(ap, RX_MINI_SIZE);
1536 else
1537 printk(KERN_ERR "%s: Someone is busy refilling "
1538 "the RX mini ring\n", ap->name);
1539 }
1540 return 0;
1541
1542 init_error:
1543 ace_init_cleanup(dev);
1544 return ecode;
1545 }
1546
1547
1548 static void ace_set_rxtx_parms(struct net_device *dev, int jumbo)
1549 {
1550 struct ace_private *ap = netdev_priv(dev);
1551 struct ace_regs __iomem *regs = ap->regs;
1552 int board_idx = ap->board_idx;
1553
1554 if (board_idx >= 0) {
1555 if (!jumbo) {
1556 if (!tx_coal_tick[board_idx])
1557 writel(DEF_TX_COAL, &regs->TuneTxCoalTicks);
1558 if (!max_tx_desc[board_idx])
1559 writel(DEF_TX_MAX_DESC, &regs->TuneMaxTxDesc);
1560 if (!rx_coal_tick[board_idx])
1561 writel(DEF_RX_COAL, &regs->TuneRxCoalTicks);
1562 if (!max_rx_desc[board_idx])
1563 writel(DEF_RX_MAX_DESC, &regs->TuneMaxRxDesc);
1564 if (!tx_ratio[board_idx])
1565 writel(DEF_TX_RATIO, &regs->TxBufRat);
1566 } else {
1567 if (!tx_coal_tick[board_idx])
1568 writel(DEF_JUMBO_TX_COAL,
1569 &regs->TuneTxCoalTicks);
1570 if (!max_tx_desc[board_idx])
1571 writel(DEF_JUMBO_TX_MAX_DESC,
1572 &regs->TuneMaxTxDesc);
1573 if (!rx_coal_tick[board_idx])
1574 writel(DEF_JUMBO_RX_COAL,
1575 &regs->TuneRxCoalTicks);
1576 if (!max_rx_desc[board_idx])
1577 writel(DEF_JUMBO_RX_MAX_DESC,
1578 &regs->TuneMaxRxDesc);
1579 if (!tx_ratio[board_idx])
1580 writel(DEF_JUMBO_TX_RATIO, &regs->TxBufRat);
1581 }
1582 }
1583 }
1584
1585
1586 static void ace_watchdog(struct net_device *data)
1587 {
1588 struct net_device *dev = data;
1589 struct ace_private *ap = netdev_priv(dev);
1590 struct ace_regs __iomem *regs = ap->regs;
1591
1592 /*
1593 * We haven't received a stats update event for more than 2.5
1594 * seconds and there is data in the transmit queue, thus we
1595 * asume the card is stuck.
1596 */
1597 if (*ap->tx_csm != ap->tx_ret_csm) {
1598 printk(KERN_WARNING "%s: Transmitter is stuck, %08x\n",
1599 dev->name, (unsigned int)readl(&regs->HostCtrl));
1600 /* This can happen due to ieee flow control. */
1601 } else {
1602 printk(KERN_DEBUG "%s: BUG... transmitter died. Kicking it.\n",
1603 dev->name);
1604 #if 0
1605 netif_wake_queue(dev);
1606 #endif
1607 }
1608 }
1609
1610
1611 static void ace_tasklet(unsigned long dev)
1612 {
1613 struct ace_private *ap = netdev_priv((struct net_device *)dev);
1614 int cur_size;
1615
1616 cur_size = atomic_read(&ap->cur_rx_bufs);
1617 if ((cur_size < RX_LOW_STD_THRES) &&
1618 !test_and_set_bit(0, &ap->std_refill_busy)) {
1619 #ifdef DEBUG
1620 printk("refilling buffers (current %i)\n", cur_size);
1621 #endif
1622 ace_load_std_rx_ring(ap, RX_RING_SIZE - cur_size);
1623 }
1624
1625 if (ap->version >= 2) {
1626 cur_size = atomic_read(&ap->cur_mini_bufs);
1627 if ((cur_size < RX_LOW_MINI_THRES) &&
1628 !test_and_set_bit(0, &ap->mini_refill_busy)) {
1629 #ifdef DEBUG
1630 printk("refilling mini buffers (current %i)\n",
1631 cur_size);
1632 #endif
1633 ace_load_mini_rx_ring(ap, RX_MINI_SIZE - cur_size);
1634 }
1635 }
1636
1637 cur_size = atomic_read(&ap->cur_jumbo_bufs);
1638 if (ap->jumbo && (cur_size < RX_LOW_JUMBO_THRES) &&
1639 !test_and_set_bit(0, &ap->jumbo_refill_busy)) {
1640 #ifdef DEBUG
1641 printk("refilling jumbo buffers (current %i)\n", cur_size);
1642 #endif
1643 ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE - cur_size);
1644 }
1645 ap->tasklet_pending = 0;
1646 }
1647
1648
1649 /*
1650 * Copy the contents of the NIC's trace buffer to kernel memory.
1651 */
1652 static void ace_dump_trace(struct ace_private *ap)
1653 {
1654 #if 0
1655 if (!ap->trace_buf)
1656 if (!(ap->trace_buf = kmalloc(ACE_TRACE_SIZE, GFP_KERNEL)))
1657 return;
1658 #endif
1659 }
1660
1661
1662 /*
1663 * Load the standard rx ring.
1664 *
1665 * Loading rings is safe without holding the spin lock since this is
1666 * done only before the device is enabled, thus no interrupts are
1667 * generated and by the interrupt handler/tasklet handler.
1668 */
1669 static void ace_load_std_rx_ring(struct ace_private *ap, int nr_bufs)
1670 {
1671 struct ace_regs __iomem *regs = ap->regs;
1672 short i, idx;
1673
1674
1675 prefetchw(&ap->cur_rx_bufs);
1676
1677 idx = ap->rx_std_skbprd;
1678
1679 for (i = 0; i < nr_bufs; i++) {
1680 struct sk_buff *skb;
1681 struct rx_desc *rd;
1682 dma_addr_t mapping;
1683
1684 skb = alloc_skb(ACE_STD_BUFSIZE + NET_IP_ALIGN, GFP_ATOMIC);
1685 if (!skb)
1686 break;
1687
1688 skb_reserve(skb, NET_IP_ALIGN);
1689 mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1690 offset_in_page(skb->data),
1691 ACE_STD_BUFSIZE,
1692 PCI_DMA_FROMDEVICE);
1693 ap->skb->rx_std_skbuff[idx].skb = skb;
1694 pci_unmap_addr_set(&ap->skb->rx_std_skbuff[idx],
1695 mapping, mapping);
1696
1697 rd = &ap->rx_std_ring[idx];
1698 set_aceaddr(&rd->addr, mapping);
1699 rd->size = ACE_STD_BUFSIZE;
1700 rd->idx = idx;
1701 idx = (idx + 1) % RX_STD_RING_ENTRIES;
1702 }
1703
1704 if (!i)
1705 goto error_out;
1706
1707 atomic_add(i, &ap->cur_rx_bufs);
1708 ap->rx_std_skbprd = idx;
1709
1710 if (ACE_IS_TIGON_I(ap)) {
1711 struct cmd cmd;
1712 cmd.evt = C_SET_RX_PRD_IDX;
1713 cmd.code = 0;
1714 cmd.idx = ap->rx_std_skbprd;
1715 ace_issue_cmd(regs, &cmd);
1716 } else {
1717 writel(idx, &regs->RxStdPrd);
1718 wmb();
1719 }
1720
1721 out:
1722 clear_bit(0, &ap->std_refill_busy);
1723 return;
1724
1725 error_out:
1726 printk(KERN_INFO "Out of memory when allocating "
1727 "standard receive buffers\n");
1728 goto out;
1729 }
1730
1731
1732 static void ace_load_mini_rx_ring(struct ace_private *ap, int nr_bufs)
1733 {
1734 struct ace_regs __iomem *regs = ap->regs;
1735 short i, idx;
1736
1737 prefetchw(&ap->cur_mini_bufs);
1738
1739 idx = ap->rx_mini_skbprd;
1740 for (i = 0; i < nr_bufs; i++) {
1741 struct sk_buff *skb;
1742 struct rx_desc *rd;
1743 dma_addr_t mapping;
1744
1745 skb = alloc_skb(ACE_MINI_BUFSIZE + NET_IP_ALIGN, GFP_ATOMIC);
1746 if (!skb)
1747 break;
1748
1749 skb_reserve(skb, NET_IP_ALIGN);
1750 mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1751 offset_in_page(skb->data),
1752 ACE_MINI_BUFSIZE,
1753 PCI_DMA_FROMDEVICE);
1754 ap->skb->rx_mini_skbuff[idx].skb = skb;
1755 pci_unmap_addr_set(&ap->skb->rx_mini_skbuff[idx],
1756 mapping, mapping);
1757
1758 rd = &ap->rx_mini_ring[idx];
1759 set_aceaddr(&rd->addr, mapping);
1760 rd->size = ACE_MINI_BUFSIZE;
1761 rd->idx = idx;
1762 idx = (idx + 1) % RX_MINI_RING_ENTRIES;
1763 }
1764
1765 if (!i)
1766 goto error_out;
1767
1768 atomic_add(i, &ap->cur_mini_bufs);
1769
1770 ap->rx_mini_skbprd = idx;
1771
1772 writel(idx, &regs->RxMiniPrd);
1773 wmb();
1774
1775 out:
1776 clear_bit(0, &ap->mini_refill_busy);
1777 return;
1778 error_out:
1779 printk(KERN_INFO "Out of memory when allocating "
1780 "mini receive buffers\n");
1781 goto out;
1782 }
1783
1784
1785 /*
1786 * Load the jumbo rx ring, this may happen at any time if the MTU
1787 * is changed to a value > 1500.
1788 */
1789 static void ace_load_jumbo_rx_ring(struct ace_private *ap, int nr_bufs)
1790 {
1791 struct ace_regs __iomem *regs = ap->regs;
1792 short i, idx;
1793
1794 idx = ap->rx_jumbo_skbprd;
1795
1796 for (i = 0; i < nr_bufs; i++) {
1797 struct sk_buff *skb;
1798 struct rx_desc *rd;
1799 dma_addr_t mapping;
1800
1801 skb = alloc_skb(ACE_JUMBO_BUFSIZE + NET_IP_ALIGN, GFP_ATOMIC);
1802 if (!skb)
1803 break;
1804
1805 skb_reserve(skb, NET_IP_ALIGN);
1806 mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
1807 offset_in_page(skb->data),
1808 ACE_JUMBO_BUFSIZE,
1809 PCI_DMA_FROMDEVICE);
1810 ap->skb->rx_jumbo_skbuff[idx].skb = skb;
1811 pci_unmap_addr_set(&ap->skb->rx_jumbo_skbuff[idx],
1812 mapping, mapping);
1813
1814 rd = &ap->rx_jumbo_ring[idx];
1815 set_aceaddr(&rd->addr, mapping);
1816 rd->size = ACE_JUMBO_BUFSIZE;
1817 rd->idx = idx;
1818 idx = (idx + 1) % RX_JUMBO_RING_ENTRIES;
1819 }
1820
1821 if (!i)
1822 goto error_out;
1823
1824 atomic_add(i, &ap->cur_jumbo_bufs);
1825 ap->rx_jumbo_skbprd = idx;
1826
1827 if (ACE_IS_TIGON_I(ap)) {
1828 struct cmd cmd;
1829 cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1830 cmd.code = 0;
1831 cmd.idx = ap->rx_jumbo_skbprd;
1832 ace_issue_cmd(regs, &cmd);
1833 } else {
1834 writel(idx, &regs->RxJumboPrd);
1835 wmb();
1836 }
1837
1838 out:
1839 clear_bit(0, &ap->jumbo_refill_busy);
1840 return;
1841 error_out:
1842 if (net_ratelimit())
1843 printk(KERN_INFO "Out of memory when allocating "
1844 "jumbo receive buffers\n");
1845 goto out;
1846 }
1847
1848
1849 /*
1850 * All events are considered to be slow (RX/TX ints do not generate
1851 * events) and are handled here, outside the main interrupt handler,
1852 * to reduce the size of the handler.
1853 */
1854 static u32 ace_handle_event(struct net_device *dev, u32 evtcsm, u32 evtprd)
1855 {
1856 struct ace_private *ap;
1857
1858 ap = netdev_priv(dev);
1859
1860 while (evtcsm != evtprd) {
1861 switch (ap->evt_ring[evtcsm].evt) {
1862 case E_FW_RUNNING:
1863 printk(KERN_INFO "%s: Firmware up and running\n",
1864 ap->name);
1865 ap->fw_running = 1;
1866 wmb();
1867 break;
1868 case E_STATS_UPDATED:
1869 break;
1870 case E_LNK_STATE:
1871 {
1872 u16 code = ap->evt_ring[evtcsm].code;
1873 switch (code) {
1874 case E_C_LINK_UP:
1875 {
1876 u32 state = readl(&ap->regs->GigLnkState);
1877 printk(KERN_WARNING "%s: Optical link UP "
1878 "(%s Duplex, Flow Control: %s%s)\n",
1879 ap->name,
1880 state & LNK_FULL_DUPLEX ? "Full":"Half",
1881 state & LNK_TX_FLOW_CTL_Y ? "TX " : "",
1882 state & LNK_RX_FLOW_CTL_Y ? "RX" : "");
1883 break;
1884 }
1885 case E_C_LINK_DOWN:
1886 printk(KERN_WARNING "%s: Optical link DOWN\n",
1887 ap->name);
1888 break;
1889 case E_C_LINK_10_100:
1890 printk(KERN_WARNING "%s: 10/100BaseT link "
1891 "UP\n", ap->name);
1892 break;
1893 default:
1894 printk(KERN_ERR "%s: Unknown optical link "
1895 "state %02x\n", ap->name, code);
1896 }
1897 break;
1898 }
1899 case E_ERROR:
1900 switch(ap->evt_ring[evtcsm].code) {
1901 case E_C_ERR_INVAL_CMD:
1902 printk(KERN_ERR "%s: invalid command error\n",
1903 ap->name);
1904 break;
1905 case E_C_ERR_UNIMP_CMD:
1906 printk(KERN_ERR "%s: unimplemented command "
1907 "error\n", ap->name);
1908 break;
1909 case E_C_ERR_BAD_CFG:
1910 printk(KERN_ERR "%s: bad config error\n",
1911 ap->name);
1912 break;
1913 default:
1914 printk(KERN_ERR "%s: unknown error %02x\n",
1915 ap->name, ap->evt_ring[evtcsm].code);
1916 }
1917 break;
1918 case E_RESET_JUMBO_RNG:
1919 {
1920 int i;
1921 for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
1922 if (ap->skb->rx_jumbo_skbuff[i].skb) {
1923 ap->rx_jumbo_ring[i].size = 0;
1924 set_aceaddr(&ap->rx_jumbo_ring[i].addr, 0);
1925 dev_kfree_skb(ap->skb->rx_jumbo_skbuff[i].skb);
1926 ap->skb->rx_jumbo_skbuff[i].skb = NULL;
1927 }
1928 }
1929
1930 if (ACE_IS_TIGON_I(ap)) {
1931 struct cmd cmd;
1932 cmd.evt = C_SET_RX_JUMBO_PRD_IDX;
1933 cmd.code = 0;
1934 cmd.idx = 0;
1935 ace_issue_cmd(ap->regs, &cmd);
1936 } else {
1937 writel(0, &((ap->regs)->RxJumboPrd));
1938 wmb();
1939 }
1940
1941 ap->jumbo = 0;
1942 ap->rx_jumbo_skbprd = 0;
1943 printk(KERN_INFO "%s: Jumbo ring flushed\n",
1944 ap->name);
1945 clear_bit(0, &ap->jumbo_refill_busy);
1946 break;
1947 }
1948 default:
1949 printk(KERN_ERR "%s: Unhandled event 0x%02x\n",
1950 ap->name, ap->evt_ring[evtcsm].evt);
1951 }
1952 evtcsm = (evtcsm + 1) % EVT_RING_ENTRIES;
1953 }
1954
1955 return evtcsm;
1956 }
1957
1958
1959 static void ace_rx_int(struct net_device *dev, u32 rxretprd, u32 rxretcsm)
1960 {
1961 struct ace_private *ap = netdev_priv(dev);
1962 u32 idx;
1963 int mini_count = 0, std_count = 0;
1964
1965 idx = rxretcsm;
1966
1967 prefetchw(&ap->cur_rx_bufs);
1968 prefetchw(&ap->cur_mini_bufs);
1969
1970 while (idx != rxretprd) {
1971 struct ring_info *rip;
1972 struct sk_buff *skb;
1973 struct rx_desc *rxdesc, *retdesc;
1974 u32 skbidx;
1975 int bd_flags, desc_type, mapsize;
1976 u16 csum;
1977
1978
1979 /* make sure the rx descriptor isn't read before rxretprd */
1980 if (idx == rxretcsm)
1981 rmb();
1982
1983 retdesc = &ap->rx_return_ring[idx];
1984 skbidx = retdesc->idx;
1985 bd_flags = retdesc->flags;
1986 desc_type = bd_flags & (BD_FLG_JUMBO | BD_FLG_MINI);
1987
1988 switch(desc_type) {
1989 /*
1990 * Normal frames do not have any flags set
1991 *
1992 * Mini and normal frames arrive frequently,
1993 * so use a local counter to avoid doing
1994 * atomic operations for each packet arriving.
1995 */
1996 case 0:
1997 rip = &ap->skb->rx_std_skbuff[skbidx];
1998 mapsize = ACE_STD_BUFSIZE;
1999 rxdesc = &ap->rx_std_ring[skbidx];
2000 std_count++;
2001 break;
2002 case BD_FLG_JUMBO:
2003 rip = &ap->skb->rx_jumbo_skbuff[skbidx];
2004 mapsize = ACE_JUMBO_BUFSIZE;
2005 rxdesc = &ap->rx_jumbo_ring[skbidx];
2006 atomic_dec(&ap->cur_jumbo_bufs);
2007 break;
2008 case BD_FLG_MINI:
2009 rip = &ap->skb->rx_mini_skbuff[skbidx];
2010 mapsize = ACE_MINI_BUFSIZE;
2011 rxdesc = &ap->rx_mini_ring[skbidx];
2012 mini_count++;
2013 break;
2014 default:
2015 printk(KERN_INFO "%s: unknown frame type (0x%02x) "
2016 "returned by NIC\n", dev->name,
2017 retdesc->flags);
2018 goto error;
2019 }
2020
2021 skb = rip->skb;
2022 rip->skb = NULL;
2023 pci_unmap_page(ap->pdev,
2024 pci_unmap_addr(rip, mapping),
2025 mapsize,
2026 PCI_DMA_FROMDEVICE);
2027 skb_put(skb, retdesc->size);
2028
2029 /*
2030 * Fly baby, fly!
2031 */
2032 csum = retdesc->tcp_udp_csum;
2033
2034 skb->dev = dev;
2035 skb->protocol = eth_type_trans(skb, dev);
2036
2037 /*
2038 * Instead of forcing the poor tigon mips cpu to calculate
2039 * pseudo hdr checksum, we do this ourselves.
2040 */
2041 if (bd_flags & BD_FLG_TCP_UDP_SUM) {
2042 skb->csum = htons(csum);
2043 skb->ip_summed = CHECKSUM_COMPLETE;
2044 } else {
2045 skb->ip_summed = CHECKSUM_NONE;
2046 }
2047
2048 /* send it up */
2049 #if ACENIC_DO_VLAN
2050 if (ap->vlgrp && (bd_flags & BD_FLG_VLAN_TAG)) {
2051 vlan_hwaccel_rx(skb, ap->vlgrp, retdesc->vlan);
2052 } else
2053 #endif
2054 netif_rx(skb);
2055
2056 dev->last_rx = jiffies;
2057 ap->stats.rx_packets++;
2058 ap->stats.rx_bytes += retdesc->size;
2059
2060 idx = (idx + 1) % RX_RETURN_RING_ENTRIES;
2061 }
2062
2063 atomic_sub(std_count, &ap->cur_rx_bufs);
2064 if (!ACE_IS_TIGON_I(ap))
2065 atomic_sub(mini_count, &ap->cur_mini_bufs);
2066
2067 out:
2068 /*
2069 * According to the documentation RxRetCsm is obsolete with
2070 * the 12.3.x Firmware - my Tigon I NICs seem to disagree!
2071 */
2072 if (ACE_IS_TIGON_I(ap)) {
2073 writel(idx, &ap->regs->RxRetCsm);
2074 }
2075 ap->cur_rx = idx;
2076
2077 return;
2078 error:
2079 idx = rxretprd;
2080 goto out;
2081 }
2082
2083
2084 static inline void ace_tx_int(struct net_device *dev,
2085 u32 txcsm, u32 idx)
2086 {
2087 struct ace_private *ap = netdev_priv(dev);
2088
2089 do {
2090 struct sk_buff *skb;
2091 dma_addr_t mapping;
2092 struct tx_ring_info *info;
2093
2094 info = ap->skb->tx_skbuff + idx;
2095 skb = info->skb;
2096 mapping = pci_unmap_addr(info, mapping);
2097
2098 if (mapping) {
2099 pci_unmap_page(ap->pdev, mapping,
2100 pci_unmap_len(info, maplen),
2101 PCI_DMA_TODEVICE);
2102 pci_unmap_addr_set(info, mapping, 0);
2103 }
2104
2105 if (skb) {
2106 ap->stats.tx_packets++;
2107 ap->stats.tx_bytes += skb->len;
2108 dev_kfree_skb_irq(skb);
2109 info->skb = NULL;
2110 }
2111
2112 idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2113 } while (idx != txcsm);
2114
2115 if (netif_queue_stopped(dev))
2116 netif_wake_queue(dev);
2117
2118 wmb();
2119 ap->tx_ret_csm = txcsm;
2120
2121 /* So... tx_ret_csm is advanced _after_ check for device wakeup.
2122 *
2123 * We could try to make it before. In this case we would get
2124 * the following race condition: hard_start_xmit on other cpu
2125 * enters after we advanced tx_ret_csm and fills space,
2126 * which we have just freed, so that we make illegal device wakeup.
2127 * There is no good way to workaround this (at entry
2128 * to ace_start_xmit detects this condition and prevents
2129 * ring corruption, but it is not a good workaround.)
2130 *
2131 * When tx_ret_csm is advanced after, we wake up device _only_
2132 * if we really have some space in ring (though the core doing
2133 * hard_start_xmit can see full ring for some period and has to
2134 * synchronize.) Superb.
2135 * BUT! We get another subtle race condition. hard_start_xmit
2136 * may think that ring is full between wakeup and advancing
2137 * tx_ret_csm and will stop device instantly! It is not so bad.
2138 * We are guaranteed that there is something in ring, so that
2139 * the next irq will resume transmission. To speedup this we could
2140 * mark descriptor, which closes ring with BD_FLG_COAL_NOW
2141 * (see ace_start_xmit).
2142 *
2143 * Well, this dilemma exists in all lock-free devices.
2144 * We, following scheme used in drivers by Donald Becker,
2145 * select the least dangerous.
2146 * --ANK
2147 */
2148 }
2149
2150
2151 static irqreturn_t ace_interrupt(int irq, void *dev_id, struct pt_regs *ptregs)
2152 {
2153 struct net_device *dev = (struct net_device *)dev_id;
2154 struct ace_private *ap = netdev_priv(dev);
2155 struct ace_regs __iomem *regs = ap->regs;
2156 u32 idx;
2157 u32 txcsm, rxretcsm, rxretprd;
2158 u32 evtcsm, evtprd;
2159
2160 /*
2161 * In case of PCI shared interrupts or spurious interrupts,
2162 * we want to make sure it is actually our interrupt before
2163 * spending any time in here.
2164 */
2165 if (!(readl(&regs->HostCtrl) & IN_INT))
2166 return IRQ_NONE;
2167
2168 /*
2169 * ACK intr now. Otherwise we will lose updates to rx_ret_prd,
2170 * which happened _after_ rxretprd = *ap->rx_ret_prd; but before
2171 * writel(0, &regs->Mb0Lo).
2172 *
2173 * "IRQ avoidance" recommended in docs applies to IRQs served
2174 * threads and it is wrong even for that case.
2175 */
2176 writel(0, &regs->Mb0Lo);
2177 readl(&regs->Mb0Lo);
2178
2179 /*
2180 * There is no conflict between transmit handling in
2181 * start_xmit and receive processing, thus there is no reason
2182 * to take a spin lock for RX handling. Wait until we start
2183 * working on the other stuff - hey we don't need a spin lock
2184 * anymore.
2185 */
2186 rxretprd = *ap->rx_ret_prd;
2187 rxretcsm = ap->cur_rx;
2188
2189 if (rxretprd != rxretcsm)
2190 ace_rx_int(dev, rxretprd, rxretcsm);
2191
2192 txcsm = *ap->tx_csm;
2193 idx = ap->tx_ret_csm;
2194
2195 if (txcsm != idx) {
2196 /*
2197 * If each skb takes only one descriptor this check degenerates
2198 * to identity, because new space has just been opened.
2199 * But if skbs are fragmented we must check that this index
2200 * update releases enough of space, otherwise we just
2201 * wait for device to make more work.
2202 */
2203 if (!tx_ring_full(ap, txcsm, ap->tx_prd))
2204 ace_tx_int(dev, txcsm, idx);
2205 }
2206
2207 evtcsm = readl(&regs->EvtCsm);
2208 evtprd = *ap->evt_prd;
2209
2210 if (evtcsm != evtprd) {
2211 evtcsm = ace_handle_event(dev, evtcsm, evtprd);
2212 writel(evtcsm, &regs->EvtCsm);
2213 }
2214
2215 /*
2216 * This has to go last in the interrupt handler and run with
2217 * the spin lock released ... what lock?
2218 */
2219 if (netif_running(dev)) {
2220 int cur_size;
2221 int run_tasklet = 0;
2222
2223 cur_size = atomic_read(&ap->cur_rx_bufs);
2224 if (cur_size < RX_LOW_STD_THRES) {
2225 if ((cur_size < RX_PANIC_STD_THRES) &&
2226 !test_and_set_bit(0, &ap->std_refill_busy)) {
2227 #ifdef DEBUG
2228 printk("low on std buffers %i\n", cur_size);
2229 #endif
2230 ace_load_std_rx_ring(ap,
2231 RX_RING_SIZE - cur_size);
2232 } else
2233 run_tasklet = 1;
2234 }
2235
2236 if (!ACE_IS_TIGON_I(ap)) {
2237 cur_size = atomic_read(&ap->cur_mini_bufs);
2238 if (cur_size < RX_LOW_MINI_THRES) {
2239 if ((cur_size < RX_PANIC_MINI_THRES) &&
2240 !test_and_set_bit(0,
2241 &ap->mini_refill_busy)) {
2242 #ifdef DEBUG
2243 printk("low on mini buffers %i\n",
2244 cur_size);
2245 #endif
2246 ace_load_mini_rx_ring(ap, RX_MINI_SIZE - cur_size);
2247 } else
2248 run_tasklet = 1;
2249 }
2250 }
2251
2252 if (ap->jumbo) {
2253 cur_size = atomic_read(&ap->cur_jumbo_bufs);
2254 if (cur_size < RX_LOW_JUMBO_THRES) {
2255 if ((cur_size < RX_PANIC_JUMBO_THRES) &&
2256 !test_and_set_bit(0,
2257 &ap->jumbo_refill_busy)){
2258 #ifdef DEBUG
2259 printk("low on jumbo buffers %i\n",
2260 cur_size);
2261 #endif
2262 ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE - cur_size);
2263 } else
2264 run_tasklet = 1;
2265 }
2266 }
2267 if (run_tasklet && !ap->tasklet_pending) {
2268 ap->tasklet_pending = 1;
2269 tasklet_schedule(&ap->ace_tasklet);
2270 }
2271 }
2272
2273 return IRQ_HANDLED;
2274 }
2275
2276
2277 #if ACENIC_DO_VLAN
2278 static void ace_vlan_rx_register(struct net_device *dev, struct vlan_group *grp)
2279 {
2280 struct ace_private *ap = netdev_priv(dev);
2281 unsigned long flags;
2282
2283 local_irq_save(flags);
2284 ace_mask_irq(dev);
2285
2286 ap->vlgrp = grp;
2287
2288 ace_unmask_irq(dev);
2289 local_irq_restore(flags);
2290 }
2291
2292
2293 static void ace_vlan_rx_kill_vid(struct net_device *dev, unsigned short vid)
2294 {
2295 struct ace_private *ap = netdev_priv(dev);
2296 unsigned long flags;
2297
2298 local_irq_save(flags);
2299 ace_mask_irq(dev);
2300
2301 if (ap->vlgrp)
2302 ap->vlgrp->vlan_devices[vid] = NULL;
2303
2304 ace_unmask_irq(dev);
2305 local_irq_restore(flags);
2306 }
2307 #endif /* ACENIC_DO_VLAN */
2308
2309
2310 static int ace_open(struct net_device *dev)
2311 {
2312 struct ace_private *ap = netdev_priv(dev);
2313 struct ace_regs __iomem *regs = ap->regs;
2314 struct cmd cmd;
2315
2316 if (!(ap->fw_running)) {
2317 printk(KERN_WARNING "%s: Firmware not running!\n", dev->name);
2318 return -EBUSY;
2319 }
2320
2321 writel(dev->mtu + ETH_HLEN + 4, &regs->IfMtu);
2322
2323 cmd.evt = C_CLEAR_STATS;
2324 cmd.code = 0;
2325 cmd.idx = 0;
2326 ace_issue_cmd(regs, &cmd);
2327
2328 cmd.evt = C_HOST_STATE;
2329 cmd.code = C_C_STACK_UP;
2330 cmd.idx = 0;
2331 ace_issue_cmd(regs, &cmd);
2332
2333 if (ap->jumbo &&
2334 !test_and_set_bit(0, &ap->jumbo_refill_busy))
2335 ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE);
2336
2337 if (dev->flags & IFF_PROMISC) {
2338 cmd.evt = C_SET_PROMISC_MODE;
2339 cmd.code = C_C_PROMISC_ENABLE;
2340 cmd.idx = 0;
2341 ace_issue_cmd(regs, &cmd);
2342
2343 ap->promisc = 1;
2344 }else
2345 ap->promisc = 0;
2346 ap->mcast_all = 0;
2347
2348 #if 0
2349 cmd.evt = C_LNK_NEGOTIATION;
2350 cmd.code = 0;
2351 cmd.idx = 0;
2352 ace_issue_cmd(regs, &cmd);
2353 #endif
2354
2355 netif_start_queue(dev);
2356
2357 /*
2358 * Setup the bottom half rx ring refill handler
2359 */
2360 tasklet_init(&ap->ace_tasklet, ace_tasklet, (unsigned long)dev);
2361 return 0;
2362 }
2363
2364
2365 static int ace_close(struct net_device *dev)
2366 {
2367 struct ace_private *ap = netdev_priv(dev);
2368 struct ace_regs __iomem *regs = ap->regs;
2369 struct cmd cmd;
2370 unsigned long flags;
2371 short i;
2372
2373 /*
2374 * Without (or before) releasing irq and stopping hardware, this
2375 * is an absolute non-sense, by the way. It will be reset instantly
2376 * by the first irq.
2377 */
2378 netif_stop_queue(dev);
2379
2380
2381 if (ap->promisc) {
2382 cmd.evt = C_SET_PROMISC_MODE;
2383 cmd.code = C_C_PROMISC_DISABLE;
2384 cmd.idx = 0;
2385 ace_issue_cmd(regs, &cmd);
2386 ap->promisc = 0;
2387 }
2388
2389 cmd.evt = C_HOST_STATE;
2390 cmd.code = C_C_STACK_DOWN;
2391 cmd.idx = 0;
2392 ace_issue_cmd(regs, &cmd);
2393
2394 tasklet_kill(&ap->ace_tasklet);
2395
2396 /*
2397 * Make sure one CPU is not processing packets while
2398 * buffers are being released by another.
2399 */
2400
2401 local_irq_save(flags);
2402 ace_mask_irq(dev);
2403
2404 for (i = 0; i < ACE_TX_RING_ENTRIES(ap); i++) {
2405 struct sk_buff *skb;
2406 dma_addr_t mapping;
2407 struct tx_ring_info *info;
2408
2409 info = ap->skb->tx_skbuff + i;
2410 skb = info->skb;
2411 mapping = pci_unmap_addr(info, mapping);
2412
2413 if (mapping) {
2414 if (ACE_IS_TIGON_I(ap)) {
2415 struct tx_desc __iomem *tx
2416 = (struct tx_desc __iomem *) &ap->tx_ring[i];
2417 writel(0, &tx->addr.addrhi);
2418 writel(0, &tx->addr.addrlo);
2419 writel(0, &tx->flagsize);
2420 } else
2421 memset(ap->tx_ring + i, 0,
2422 sizeof(struct tx_desc));
2423 pci_unmap_page(ap->pdev, mapping,
2424 pci_unmap_len(info, maplen),
2425 PCI_DMA_TODEVICE);
2426 pci_unmap_addr_set(info, mapping, 0);
2427 }
2428 if (skb) {
2429 dev_kfree_skb(skb);
2430 info->skb = NULL;
2431 }
2432 }
2433
2434 if (ap->jumbo) {
2435 cmd.evt = C_RESET_JUMBO_RNG;
2436 cmd.code = 0;
2437 cmd.idx = 0;
2438 ace_issue_cmd(regs, &cmd);
2439 }
2440
2441 ace_unmask_irq(dev);
2442 local_irq_restore(flags);
2443
2444 return 0;
2445 }
2446
2447
2448 static inline dma_addr_t
2449 ace_map_tx_skb(struct ace_private *ap, struct sk_buff *skb,
2450 struct sk_buff *tail, u32 idx)
2451 {
2452 dma_addr_t mapping;
2453 struct tx_ring_info *info;
2454
2455 mapping = pci_map_page(ap->pdev, virt_to_page(skb->data),
2456 offset_in_page(skb->data),
2457 skb->len, PCI_DMA_TODEVICE);
2458
2459 info = ap->skb->tx_skbuff + idx;
2460 info->skb = tail;
2461 pci_unmap_addr_set(info, mapping, mapping);
2462 pci_unmap_len_set(info, maplen, skb->len);
2463 return mapping;
2464 }
2465
2466
2467 static inline void
2468 ace_load_tx_bd(struct ace_private *ap, struct tx_desc *desc, u64 addr,
2469 u32 flagsize, u32 vlan_tag)
2470 {
2471 #if !USE_TX_COAL_NOW
2472 flagsize &= ~BD_FLG_COAL_NOW;
2473 #endif
2474
2475 if (ACE_IS_TIGON_I(ap)) {
2476 struct tx_desc __iomem *io = (struct tx_desc __iomem *) desc;
2477 writel(addr >> 32, &io->addr.addrhi);
2478 writel(addr & 0xffffffff, &io->addr.addrlo);
2479 writel(flagsize, &io->flagsize);
2480 #if ACENIC_DO_VLAN
2481 writel(vlan_tag, &io->vlanres);
2482 #endif
2483 } else {
2484 desc->addr.addrhi = addr >> 32;
2485 desc->addr.addrlo = addr;
2486 desc->flagsize = flagsize;
2487 #if ACENIC_DO_VLAN
2488 desc->vlanres = vlan_tag;
2489 #endif
2490 }
2491 }
2492
2493
2494 static int ace_start_xmit(struct sk_buff *skb, struct net_device *dev)
2495 {
2496 struct ace_private *ap = netdev_priv(dev);
2497 struct ace_regs __iomem *regs = ap->regs;
2498 struct tx_desc *desc;
2499 u32 idx, flagsize;
2500 unsigned long maxjiff = jiffies + 3*HZ;
2501
2502 restart:
2503 idx = ap->tx_prd;
2504
2505 if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2506 goto overflow;
2507
2508 if (!skb_shinfo(skb)->nr_frags) {
2509 dma_addr_t mapping;
2510 u32 vlan_tag = 0;
2511
2512 mapping = ace_map_tx_skb(ap, skb, skb, idx);
2513 flagsize = (skb->len << 16) | (BD_FLG_END);
2514 if (skb->ip_summed == CHECKSUM_PARTIAL)
2515 flagsize |= BD_FLG_TCP_UDP_SUM;
2516 #if ACENIC_DO_VLAN
2517 if (vlan_tx_tag_present(skb)) {
2518 flagsize |= BD_FLG_VLAN_TAG;
2519 vlan_tag = vlan_tx_tag_get(skb);
2520 }
2521 #endif
2522 desc = ap->tx_ring + idx;
2523 idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2524
2525 /* Look at ace_tx_int for explanations. */
2526 if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2527 flagsize |= BD_FLG_COAL_NOW;
2528
2529 ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2530 } else {
2531 dma_addr_t mapping;
2532 u32 vlan_tag = 0;
2533 int i, len = 0;
2534
2535 mapping = ace_map_tx_skb(ap, skb, NULL, idx);
2536 flagsize = (skb_headlen(skb) << 16);
2537 if (skb->ip_summed == CHECKSUM_PARTIAL)
2538 flagsize |= BD_FLG_TCP_UDP_SUM;
2539 #if ACENIC_DO_VLAN
2540 if (vlan_tx_tag_present(skb)) {
2541 flagsize |= BD_FLG_VLAN_TAG;
2542 vlan_tag = vlan_tx_tag_get(skb);
2543 }
2544 #endif
2545
2546 ace_load_tx_bd(ap, ap->tx_ring + idx, mapping, flagsize, vlan_tag);
2547
2548 idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2549
2550 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2551 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2552 struct tx_ring_info *info;
2553
2554 len += frag->size;
2555 info = ap->skb->tx_skbuff + idx;
2556 desc = ap->tx_ring + idx;
2557
2558 mapping = pci_map_page(ap->pdev, frag->page,
2559 frag->page_offset, frag->size,
2560 PCI_DMA_TODEVICE);
2561
2562 flagsize = (frag->size << 16);
2563 if (skb->ip_summed == CHECKSUM_PARTIAL)
2564 flagsize |= BD_FLG_TCP_UDP_SUM;
2565 idx = (idx + 1) % ACE_TX_RING_ENTRIES(ap);
2566
2567 if (i == skb_shinfo(skb)->nr_frags - 1) {
2568 flagsize |= BD_FLG_END;
2569 if (tx_ring_full(ap, ap->tx_ret_csm, idx))
2570 flagsize |= BD_FLG_COAL_NOW;
2571
2572 /*
2573 * Only the last fragment frees
2574 * the skb!
2575 */
2576 info->skb = skb;
2577 } else {
2578 info->skb = NULL;
2579 }
2580 pci_unmap_addr_set(info, mapping, mapping);
2581 pci_unmap_len_set(info, maplen, frag->size);
2582 ace_load_tx_bd(ap, desc, mapping, flagsize, vlan_tag);
2583 }
2584 }
2585
2586 wmb();
2587 ap->tx_prd = idx;
2588 ace_set_txprd(regs, ap, idx);
2589
2590 if (flagsize & BD_FLG_COAL_NOW) {
2591 netif_stop_queue(dev);
2592
2593 /*
2594 * A TX-descriptor producer (an IRQ) might have gotten
2595 * inbetween, making the ring free again. Since xmit is
2596 * serialized, this is the only situation we have to
2597 * re-test.
2598 */
2599 if (!tx_ring_full(ap, ap->tx_ret_csm, idx))
2600 netif_wake_queue(dev);
2601 }
2602
2603 dev->trans_start = jiffies;
2604 return NETDEV_TX_OK;
2605
2606 overflow:
2607 /*
2608 * This race condition is unavoidable with lock-free drivers.
2609 * We wake up the queue _before_ tx_prd is advanced, so that we can
2610 * enter hard_start_xmit too early, while tx ring still looks closed.
2611 * This happens ~1-4 times per 100000 packets, so that we can allow
2612 * to loop syncing to other CPU. Probably, we need an additional
2613 * wmb() in ace_tx_intr as well.
2614 *
2615 * Note that this race is relieved by reserving one more entry
2616 * in tx ring than it is necessary (see original non-SG driver).
2617 * However, with SG we need to reserve 2*MAX_SKB_FRAGS+1, which
2618 * is already overkill.
2619 *
2620 * Alternative is to return with 1 not throttling queue. In this
2621 * case loop becomes longer, no more useful effects.
2622 */
2623 if (time_before(jiffies, maxjiff)) {
2624 barrier();
2625 cpu_relax();
2626 goto restart;
2627 }
2628
2629 /* The ring is stuck full. */
2630 printk(KERN_WARNING "%s: Transmit ring stuck full\n", dev->name);
2631 return NETDEV_TX_BUSY;
2632 }
2633
2634
2635 static int ace_change_mtu(struct net_device *dev, int new_mtu)
2636 {
2637 struct ace_private *ap = netdev_priv(dev);
2638 struct ace_regs __iomem *regs = ap->regs;
2639
2640 if (new_mtu > ACE_JUMBO_MTU)
2641 return -EINVAL;
2642
2643 writel(new_mtu + ETH_HLEN + 4, &regs->IfMtu);
2644 dev->mtu = new_mtu;
2645
2646 if (new_mtu > ACE_STD_MTU) {
2647 if (!(ap->jumbo)) {
2648 printk(KERN_INFO "%s: Enabling Jumbo frame "
2649 "support\n", dev->name);
2650 ap->jumbo = 1;
2651 if (!test_and_set_bit(0, &ap->jumbo_refill_busy))
2652 ace_load_jumbo_rx_ring(ap, RX_JUMBO_SIZE);
2653 ace_set_rxtx_parms(dev, 1);
2654 }
2655 } else {
2656 while (test_and_set_bit(0, &ap->jumbo_refill_busy));
2657 ace_sync_irq(dev->irq);
2658 ace_set_rxtx_parms(dev, 0);
2659 if (ap->jumbo) {
2660 struct cmd cmd;
2661
2662 cmd.evt = C_RESET_JUMBO_RNG;
2663 cmd.code = 0;
2664 cmd.idx = 0;
2665 ace_issue_cmd(regs, &cmd);
2666 }
2667 }
2668
2669 return 0;
2670 }
2671
2672 static int ace_get_settings(struct net_device *dev, struct ethtool_cmd *ecmd)
2673 {
2674 struct ace_private *ap = netdev_priv(dev);
2675 struct ace_regs __iomem *regs = ap->regs;
2676 u32 link;
2677
2678 memset(ecmd, 0, sizeof(struct ethtool_cmd));
2679 ecmd->supported =
2680 (SUPPORTED_10baseT_Half | SUPPORTED_10baseT_Full |
2681 SUPPORTED_100baseT_Half | SUPPORTED_100baseT_Full |
2682 SUPPORTED_1000baseT_Half | SUPPORTED_1000baseT_Full |
2683 SUPPORTED_Autoneg | SUPPORTED_FIBRE);
2684
2685 ecmd->port = PORT_FIBRE;
2686 ecmd->transceiver = XCVR_INTERNAL;
2687
2688 link = readl(&regs->GigLnkState);
2689 if (link & LNK_1000MB)
2690 ecmd->speed = SPEED_1000;
2691 else {
2692 link = readl(&regs->FastLnkState);
2693 if (link & LNK_100MB)
2694 ecmd->speed = SPEED_100;
2695 else if (link & LNK_10MB)
2696 ecmd->speed = SPEED_10;
2697 else
2698 ecmd->speed = 0;
2699 }
2700 if (link & LNK_FULL_DUPLEX)
2701 ecmd->duplex = DUPLEX_FULL;
2702 else
2703 ecmd->duplex = DUPLEX_HALF;
2704
2705 if (link & LNK_NEGOTIATE)
2706 ecmd->autoneg = AUTONEG_ENABLE;
2707 else
2708 ecmd->autoneg = AUTONEG_DISABLE;
2709
2710 #if 0
2711 /*
2712 * Current struct ethtool_cmd is insufficient
2713 */
2714 ecmd->trace = readl(&regs->TuneTrace);
2715
2716 ecmd->txcoal = readl(&regs->TuneTxCoalTicks);
2717 ecmd->rxcoal = readl(&regs->TuneRxCoalTicks);
2718 #endif
2719 ecmd->maxtxpkt = readl(&regs->TuneMaxTxDesc);
2720 ecmd->maxrxpkt = readl(&regs->TuneMaxRxDesc);
2721
2722 return 0;
2723 }
2724
2725 static int ace_set_settings(struct net_device *dev, struct ethtool_cmd *ecmd)
2726 {
2727 struct ace_private *ap = netdev_priv(dev);
2728 struct ace_regs __iomem *regs = ap->regs;
2729 u32 link, speed;
2730
2731 link = readl(&regs->GigLnkState);
2732 if (link & LNK_1000MB)
2733 speed = SPEED_1000;
2734 else {
2735 link = readl(&regs->FastLnkState);
2736 if (link & LNK_100MB)
2737 speed = SPEED_100;
2738 else if (link & LNK_10MB)
2739 speed = SPEED_10;
2740 else
2741 speed = SPEED_100;
2742 }
2743
2744 link = LNK_ENABLE | LNK_1000MB | LNK_100MB | LNK_10MB |
2745 LNK_RX_FLOW_CTL_Y | LNK_NEG_FCTL;
2746 if (!ACE_IS_TIGON_I(ap))
2747 link |= LNK_TX_FLOW_CTL_Y;
2748 if (ecmd->autoneg == AUTONEG_ENABLE)
2749 link |= LNK_NEGOTIATE;
2750 if (ecmd->speed != speed) {
2751 link &= ~(LNK_1000MB | LNK_100MB | LNK_10MB);
2752 switch (speed) {
2753 case SPEED_1000:
2754 link |= LNK_1000MB;
2755 break;
2756 case SPEED_100:
2757 link |= LNK_100MB;
2758 break;
2759 case SPEED_10:
2760 link |= LNK_10MB;
2761 break;
2762 }
2763 }
2764
2765 if (ecmd->duplex == DUPLEX_FULL)
2766 link |= LNK_FULL_DUPLEX;
2767
2768 if (link != ap->link) {
2769 struct cmd cmd;
2770 printk(KERN_INFO "%s: Renegotiating link state\n",
2771 dev->name);
2772
2773 ap->link = link;
2774 writel(link, &regs->TuneLink);
2775 if (!ACE_IS_TIGON_I(ap))
2776 writel(link, &regs->TuneFastLink);
2777 wmb();
2778
2779 cmd.evt = C_LNK_NEGOTIATION;
2780 cmd.code = 0;
2781 cmd.idx = 0;
2782 ace_issue_cmd(regs, &cmd);
2783 }
2784 return 0;
2785 }
2786
2787 static void ace_get_drvinfo(struct net_device *dev,
2788 struct ethtool_drvinfo *info)
2789 {
2790 struct ace_private *ap = netdev_priv(dev);
2791
2792 strlcpy(info->driver, "acenic", sizeof(info->driver));
2793 snprintf(info->version, sizeof(info->version), "%i.%i.%i",
2794 tigonFwReleaseMajor, tigonFwReleaseMinor,
2795 tigonFwReleaseFix);
2796
2797 if (ap->pdev)
2798 strlcpy(info->bus_info, pci_name(ap->pdev),
2799 sizeof(info->bus_info));
2800
2801 }
2802
2803 /*
2804 * Set the hardware MAC address.
2805 */
2806 static int ace_set_mac_addr(struct net_device *dev, void *p)
2807 {
2808 struct ace_private *ap = netdev_priv(dev);
2809 struct ace_regs __iomem *regs = ap->regs;
2810 struct sockaddr *addr=p;
2811 u8 *da;
2812 struct cmd cmd;
2813
2814 if(netif_running(dev))
2815 return -EBUSY;
2816
2817 memcpy(dev->dev_addr, addr->sa_data,dev->addr_len);
2818
2819 da = (u8 *)dev->dev_addr;
2820
2821 writel(da[0] << 8 | da[1], &regs->MacAddrHi);
2822 writel((da[2] << 24) | (da[3] << 16) | (da[4] << 8) | da[5],
2823 &regs->MacAddrLo);
2824
2825 cmd.evt = C_SET_MAC_ADDR;
2826 cmd.code = 0;
2827 cmd.idx = 0;
2828 ace_issue_cmd(regs, &cmd);
2829
2830 return 0;
2831 }
2832
2833
2834 static void ace_set_multicast_list(struct net_device *dev)
2835 {
2836 struct ace_private *ap = netdev_priv(dev);
2837 struct ace_regs __iomem *regs = ap->regs;
2838 struct cmd cmd;
2839
2840 if ((dev->flags & IFF_ALLMULTI) && !(ap->mcast_all)) {
2841 cmd.evt = C_SET_MULTICAST_MODE;
2842 cmd.code = C_C_MCAST_ENABLE;
2843 cmd.idx = 0;
2844 ace_issue_cmd(regs, &cmd);
2845 ap->mcast_all = 1;
2846 } else if (ap->mcast_all) {
2847 cmd.evt = C_SET_MULTICAST_MODE;
2848 cmd.code = C_C_MCAST_DISABLE;
2849 cmd.idx = 0;
2850 ace_issue_cmd(regs, &cmd);
2851 ap->mcast_all = 0;
2852 }
2853
2854 if ((dev->flags & IFF_PROMISC) && !(ap->promisc)) {
2855 cmd.evt = C_SET_PROMISC_MODE;
2856 cmd.code = C_C_PROMISC_ENABLE;
2857 cmd.idx = 0;
2858 ace_issue_cmd(regs, &cmd);
2859 ap->promisc = 1;
2860 }else if (!(dev->flags & IFF_PROMISC) && (ap->promisc)) {
2861 cmd.evt = C_SET_PROMISC_MODE;
2862 cmd.code = C_C_PROMISC_DISABLE;
2863 cmd.idx = 0;
2864 ace_issue_cmd(regs, &cmd);
2865 ap->promisc = 0;
2866 }
2867
2868 /*
2869 * For the time being multicast relies on the upper layers
2870 * filtering it properly. The Firmware does not allow one to
2871 * set the entire multicast list at a time and keeping track of
2872 * it here is going to be messy.
2873 */
2874 if ((dev->mc_count) && !(ap->mcast_all)) {
2875 cmd.evt = C_SET_MULTICAST_MODE;
2876 cmd.code = C_C_MCAST_ENABLE;
2877 cmd.idx = 0;
2878 ace_issue_cmd(regs, &cmd);
2879 }else if (!ap->mcast_all) {
2880 cmd.evt = C_SET_MULTICAST_MODE;
2881 cmd.code = C_C_MCAST_DISABLE;
2882 cmd.idx = 0;
2883 ace_issue_cmd(regs, &cmd);
2884 }
2885 }
2886
2887
2888 static struct net_device_stats *ace_get_stats(struct net_device *dev)
2889 {
2890 struct ace_private *ap = netdev_priv(dev);
2891 struct ace_mac_stats __iomem *mac_stats =
2892 (struct ace_mac_stats __iomem *)ap->regs->Stats;
2893
2894 ap->stats.rx_missed_errors = readl(&mac_stats->drop_space);
2895 ap->stats.multicast = readl(&mac_stats->kept_mc);
2896 ap->stats.collisions = readl(&mac_stats->coll);
2897
2898 return &ap->stats;
2899 }
2900
2901
2902 static void __devinit ace_copy(struct ace_regs __iomem *regs, void *src,
2903 u32 dest, int size)
2904 {
2905 void __iomem *tdest;
2906 u32 *wsrc;
2907 short tsize, i;
2908
2909 if (size <= 0)
2910 return;
2911
2912 while (size > 0) {
2913 tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2914 min_t(u32, size, ACE_WINDOW_SIZE));
2915 tdest = (void __iomem *) &regs->Window +
2916 (dest & (ACE_WINDOW_SIZE - 1));
2917 writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2918 /*
2919 * This requires byte swapping on big endian, however
2920 * writel does that for us
2921 */
2922 wsrc = src;
2923 for (i = 0; i < (tsize / 4); i++) {
2924 writel(wsrc[i], tdest + i*4);
2925 }
2926 dest += tsize;
2927 src += tsize;
2928 size -= tsize;
2929 }
2930
2931 return;
2932 }
2933
2934
2935 static void __devinit ace_clear(struct ace_regs __iomem *regs, u32 dest, int size)
2936 {
2937 void __iomem *tdest;
2938 short tsize = 0, i;
2939
2940 if (size <= 0)
2941 return;
2942
2943 while (size > 0) {
2944 tsize = min_t(u32, ((~dest & (ACE_WINDOW_SIZE - 1)) + 1),
2945 min_t(u32, size, ACE_WINDOW_SIZE));
2946 tdest = (void __iomem *) &regs->Window +
2947 (dest & (ACE_WINDOW_SIZE - 1));
2948 writel(dest & ~(ACE_WINDOW_SIZE - 1), &regs->WinBase);
2949
2950 for (i = 0; i < (tsize / 4); i++) {
2951 writel(0, tdest + i*4);
2952 }
2953
2954 dest += tsize;
2955 size -= tsize;
2956 }
2957
2958 return;
2959 }
2960
2961
2962 /*
2963 * Download the firmware into the SRAM on the NIC
2964 *
2965 * This operation requires the NIC to be halted and is performed with
2966 * interrupts disabled and with the spinlock hold.
2967 */
2968 int __devinit ace_load_firmware(struct net_device *dev)
2969 {
2970 struct ace_private *ap = netdev_priv(dev);
2971 struct ace_regs __iomem *regs = ap->regs;
2972
2973 if (!(readl(&regs->CpuCtrl) & CPU_HALTED)) {
2974 printk(KERN_ERR "%s: trying to download firmware while the "
2975 "CPU is running!\n", ap->name);
2976 return -EFAULT;
2977 }
2978
2979 /*
2980 * Do not try to clear more than 512KB or we end up seeing
2981 * funny things on NICs with only 512KB SRAM
2982 */
2983 ace_clear(regs, 0x2000, 0x80000-0x2000);
2984 if (ACE_IS_TIGON_I(ap)) {
2985 ace_copy(regs, tigonFwText, tigonFwTextAddr, tigonFwTextLen);
2986 ace_copy(regs, tigonFwData, tigonFwDataAddr, tigonFwDataLen);
2987 ace_copy(regs, tigonFwRodata, tigonFwRodataAddr,
2988 tigonFwRodataLen);
2989 ace_clear(regs, tigonFwBssAddr, tigonFwBssLen);
2990 ace_clear(regs, tigonFwSbssAddr, tigonFwSbssLen);
2991 }else if (ap->version == 2) {
2992 ace_clear(regs, tigon2FwBssAddr, tigon2FwBssLen);
2993 ace_clear(regs, tigon2FwSbssAddr, tigon2FwSbssLen);
2994 ace_copy(regs, tigon2FwText, tigon2FwTextAddr,tigon2FwTextLen);
2995 ace_copy(regs, tigon2FwRodata, tigon2FwRodataAddr,
2996 tigon2FwRodataLen);
2997 ace_copy(regs, tigon2FwData, tigon2FwDataAddr,tigon2FwDataLen);
2998 }
2999
3000 return 0;
3001 }
3002
3003
3004 /*
3005 * The eeprom on the AceNIC is an Atmel i2c EEPROM.
3006 *
3007 * Accessing the EEPROM is `interesting' to say the least - don't read
3008 * this code right after dinner.
3009 *
3010 * This is all about black magic and bit-banging the device .... I
3011 * wonder in what hospital they have put the guy who designed the i2c
3012 * specs.
3013 *
3014 * Oh yes, this is only the beginning!
3015 *
3016 * Thanks to Stevarino Webinski for helping tracking down the bugs in the
3017 * code i2c readout code by beta testing all my hacks.
3018 */
3019 static void __devinit eeprom_start(struct ace_regs __iomem *regs)
3020 {
3021 u32 local;
3022
3023 readl(&regs->LocalCtrl);
3024 udelay(ACE_SHORT_DELAY);
3025 local = readl(&regs->LocalCtrl);
3026 local |= EEPROM_DATA_OUT | EEPROM_WRITE_ENABLE;
3027 writel(local, &regs->LocalCtrl);
3028 readl(&regs->LocalCtrl);
3029 mb();
3030 udelay(ACE_SHORT_DELAY);
3031 local |= EEPROM_CLK_OUT;
3032 writel(local, &regs->LocalCtrl);
3033 readl(&regs->LocalCtrl);
3034 mb();
3035 udelay(ACE_SHORT_DELAY);
3036 local &= ~EEPROM_DATA_OUT;
3037 writel(local, &regs->LocalCtrl);
3038 readl(&regs->LocalCtrl);
3039 mb();
3040 udelay(ACE_SHORT_DELAY);
3041 local &= ~EEPROM_CLK_OUT;
3042 writel(local, &regs->LocalCtrl);
3043 readl(&regs->LocalCtrl);
3044 mb();
3045 }
3046
3047
3048 static void __devinit eeprom_prep(struct ace_regs __iomem *regs, u8 magic)
3049 {
3050 short i;
3051 u32 local;
3052
3053 udelay(ACE_SHORT_DELAY);
3054 local = readl(&regs->LocalCtrl);
3055 local &= ~EEPROM_DATA_OUT;
3056 local |= EEPROM_WRITE_ENABLE;
3057 writel(local, &regs->LocalCtrl);
3058 readl(&regs->LocalCtrl);
3059 mb();
3060
3061 for (i = 0; i < 8; i++, magic <<= 1) {
3062 udelay(ACE_SHORT_DELAY);
3063 if (magic & 0x80)
3064 local |= EEPROM_DATA_OUT;
3065 else
3066 local &= ~EEPROM_DATA_OUT;
3067 writel(local, &regs->LocalCtrl);
3068 readl(&regs->LocalCtrl);
3069 mb();
3070
3071 udelay(ACE_SHORT_DELAY);
3072 local |= EEPROM_CLK_OUT;
3073 writel(local, &regs->LocalCtrl);
3074 readl(&regs->LocalCtrl);
3075 mb();
3076 udelay(ACE_SHORT_DELAY);
3077 local &= ~(EEPROM_CLK_OUT | EEPROM_DATA_OUT);
3078 writel(local, &regs->LocalCtrl);
3079 readl(&regs->LocalCtrl);
3080 mb();
3081 }
3082 }
3083
3084
3085 static int __devinit eeprom_check_ack(struct ace_regs __iomem *regs)
3086 {
3087 int state;
3088 u32 local;
3089
3090 local = readl(&regs->LocalCtrl);
3091 local &= ~EEPROM_WRITE_ENABLE;
3092 writel(local, &regs->LocalCtrl);
3093 readl(&regs->LocalCtrl);
3094 mb();
3095 udelay(ACE_LONG_DELAY);
3096 local |= EEPROM_CLK_OUT;
3097 writel(local, &regs->LocalCtrl);
3098 readl(&regs->LocalCtrl);
3099 mb();
3100 udelay(ACE_SHORT_DELAY);
3101 /* sample data in middle of high clk */
3102 state = (readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0;
3103 udelay(ACE_SHORT_DELAY);
3104 mb();
3105 writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3106 readl(&regs->LocalCtrl);
3107 mb();
3108
3109 return state;
3110 }
3111
3112
3113 static void __devinit eeprom_stop(struct ace_regs __iomem *regs)
3114 {
3115 u32 local;
3116
3117 udelay(ACE_SHORT_DELAY);
3118 local = readl(&regs->LocalCtrl);
3119 local |= EEPROM_WRITE_ENABLE;
3120 writel(local, &regs->LocalCtrl);
3121 readl(&regs->LocalCtrl);
3122 mb();
3123 udelay(ACE_SHORT_DELAY);
3124 local &= ~EEPROM_DATA_OUT;
3125 writel(local, &regs->LocalCtrl);
3126 readl(&regs->LocalCtrl);
3127 mb();
3128 udelay(ACE_SHORT_DELAY);
3129 local |= EEPROM_CLK_OUT;
3130 writel(local, &regs->LocalCtrl);
3131 readl(&regs->LocalCtrl);
3132 mb();
3133 udelay(ACE_SHORT_DELAY);
3134 local |= EEPROM_DATA_OUT;
3135 writel(local, &regs->LocalCtrl);
3136 readl(&regs->LocalCtrl);
3137 mb();
3138 udelay(ACE_LONG_DELAY);
3139 local &= ~EEPROM_CLK_OUT;
3140 writel(local, &regs->LocalCtrl);
3141 mb();
3142 }
3143
3144
3145 /*
3146 * Read a whole byte from the EEPROM.
3147 */
3148 static int __devinit read_eeprom_byte(struct net_device *dev,
3149 unsigned long offset)
3150 {
3151 struct ace_private *ap = netdev_priv(dev);
3152 struct ace_regs __iomem *regs = ap->regs;
3153 unsigned long flags;
3154 u32 local;
3155 int result = 0;
3156 short i;
3157
3158 if (!dev) {
3159 printk(KERN_ERR "No device!\n");
3160 result = -ENODEV;
3161 goto out;
3162 }
3163
3164 /*
3165 * Don't take interrupts on this CPU will bit banging
3166 * the %#%#@$ I2C device
3167 */
3168 local_irq_save(flags);
3169
3170 eeprom_start(regs);
3171
3172 eeprom_prep(regs, EEPROM_WRITE_SELECT);
3173 if (eeprom_check_ack(regs)) {
3174 local_irq_restore(flags);
3175 printk(KERN_ERR "%s: Unable to sync eeprom\n", ap->name);
3176 result = -EIO;
3177 goto eeprom_read_error;
3178 }
3179
3180 eeprom_prep(regs, (offset >> 8) & 0xff);
3181 if (eeprom_check_ack(regs)) {
3182 local_irq_restore(flags);
3183 printk(KERN_ERR "%s: Unable to set address byte 0\n",
3184 ap->name);
3185 result = -EIO;
3186 goto eeprom_read_error;
3187 }
3188
3189 eeprom_prep(regs, offset & 0xff);
3190 if (eeprom_check_ack(regs)) {
3191 local_irq_restore(flags);
3192 printk(KERN_ERR "%s: Unable to set address byte 1\n",
3193 ap->name);
3194 result = -EIO;
3195 goto eeprom_read_error;
3196 }
3197
3198 eeprom_start(regs);
3199 eeprom_prep(regs, EEPROM_READ_SELECT);
3200 if (eeprom_check_ack(regs)) {
3201 local_irq_restore(flags);
3202 printk(KERN_ERR "%s: Unable to set READ_SELECT\n",
3203 ap->name);
3204 result = -EIO;
3205 goto eeprom_read_error;
3206 }
3207
3208 for (i = 0; i < 8; i++) {
3209 local = readl(&regs->LocalCtrl);
3210 local &= ~EEPROM_WRITE_ENABLE;
3211 writel(local, &regs->LocalCtrl);
3212 readl(&regs->LocalCtrl);
3213 udelay(ACE_LONG_DELAY);
3214 mb();
3215 local |= EEPROM_CLK_OUT;
3216 writel(local, &regs->LocalCtrl);
3217 readl(&regs->LocalCtrl);
3218 mb();
3219 udelay(ACE_SHORT_DELAY);
3220 /* sample data mid high clk */
3221 result = (result << 1) |
3222 ((readl(&regs->LocalCtrl) & EEPROM_DATA_IN) != 0);
3223 udelay(ACE_SHORT_DELAY);
3224 mb();
3225 local = readl(&regs->LocalCtrl);
3226 local &= ~EEPROM_CLK_OUT;
3227 writel(local, &regs->LocalCtrl);
3228 readl(&regs->LocalCtrl);
3229 udelay(ACE_SHORT_DELAY);
3230 mb();
3231 if (i == 7) {
3232 local |= EEPROM_WRITE_ENABLE;
3233 writel(local, &regs->LocalCtrl);
3234 readl(&regs->LocalCtrl);
3235 mb();
3236 udelay(ACE_SHORT_DELAY);
3237 }
3238 }
3239
3240 local |= EEPROM_DATA_OUT;
3241 writel(local, &regs->LocalCtrl);
3242 readl(&regs->LocalCtrl);
3243 mb();
3244 udelay(ACE_SHORT_DELAY);
3245 writel(readl(&regs->LocalCtrl) | EEPROM_CLK_OUT, &regs->LocalCtrl);
3246 readl(&regs->LocalCtrl);
3247 udelay(ACE_LONG_DELAY);
3248 writel(readl(&regs->LocalCtrl) & ~EEPROM_CLK_OUT, &regs->LocalCtrl);
3249 readl(&regs->LocalCtrl);
3250 mb();
3251 udelay(ACE_SHORT_DELAY);
3252 eeprom_stop(regs);
3253
3254 local_irq_restore(flags);
3255 out:
3256 return result;
3257
3258 eeprom_read_error:
3259 printk(KERN_ERR "%s: Unable to read eeprom byte 0x%02lx\n",
3260 ap->name, offset);
3261 goto out;
3262 }
3263
3264
3265 /*
3266 * Local variables:
3267 * compile-command: "gcc -D__SMP__ -D__KERNEL__ -DMODULE -I../../include -Wall -Wstrict-prototypes -O2 -fomit-frame-pointer -pipe -fno-strength-reduce -DMODVERSIONS -include ../../include/linux/modversions.h -c -o acenic.o acenic.c"
3268 * End:
3269 */