2 * acenic.c: Linux driver for the Alteon AceNIC Gigabit Ethernet card
3 * and other Tigon based cards.
5 * Copyright 1998-2002 by Jes Sorensen, <jes@trained-monkey.org>.
7 * Thanks to Alteon and 3Com for providing hardware and documentation
8 * enabling me to write this driver.
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.
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.
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
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
50 * Grant Grundler <grundler@cup.hp.com>: PCI write posting fixes.
53 #include <linux/module.h>
54 #include <linux/moduleparam.h>
55 #include <linux/types.h>
56 #include <linux/errno.h>
57 #include <linux/ioport.h>
58 #include <linux/pci.h>
59 #include <linux/dma-mapping.h>
60 #include <linux/kernel.h>
61 #include <linux/netdevice.h>
62 #include <linux/etherdevice.h>
63 #include <linux/skbuff.h>
64 #include <linux/init.h>
65 #include <linux/delay.h>
67 #include <linux/highmem.h>
68 #include <linux/sockios.h>
69 #include <linux/firmware.h>
70 #include <linux/slab.h>
71 #include <linux/prefetch.h>
72 #include <linux/if_vlan.h>
75 #include <linux/ethtool.h>
81 #include <asm/system.h>
84 #include <asm/byteorder.h>
85 #include <asm/uaccess.h>
88 #define DRV_NAME "acenic"
92 #ifdef CONFIG_ACENIC_OMIT_TIGON_I
93 #define ACE_IS_TIGON_I(ap) 0
94 #define ACE_TX_RING_ENTRIES(ap) MAX_TX_RING_ENTRIES
96 #define ACE_IS_TIGON_I(ap) (ap->version == 1)
97 #define ACE_TX_RING_ENTRIES(ap) ap->tx_ring_entries
100 #ifndef PCI_VENDOR_ID_ALTEON
101 #define PCI_VENDOR_ID_ALTEON 0x12ae
103 #ifndef PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE
104 #define PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE 0x0001
105 #define PCI_DEVICE_ID_ALTEON_ACENIC_COPPER 0x0002
107 #ifndef PCI_DEVICE_ID_3COM_3C985
108 #define PCI_DEVICE_ID_3COM_3C985 0x0001
110 #ifndef PCI_VENDOR_ID_NETGEAR
111 #define PCI_VENDOR_ID_NETGEAR 0x1385
112 #define PCI_DEVICE_ID_NETGEAR_GA620 0x620a
114 #ifndef PCI_DEVICE_ID_NETGEAR_GA620T
115 #define PCI_DEVICE_ID_NETGEAR_GA620T 0x630a
120 * Farallon used the DEC vendor ID by mistake and they seem not
123 #ifndef PCI_DEVICE_ID_FARALLON_PN9000SX
124 #define PCI_DEVICE_ID_FARALLON_PN9000SX 0x1a
126 #ifndef PCI_DEVICE_ID_FARALLON_PN9100T
127 #define PCI_DEVICE_ID_FARALLON_PN9100T 0xfa
129 #ifndef PCI_VENDOR_ID_SGI
130 #define PCI_VENDOR_ID_SGI 0x10a9
132 #ifndef PCI_DEVICE_ID_SGI_ACENIC
133 #define PCI_DEVICE_ID_SGI_ACENIC 0x0009
136 static DEFINE_PCI_DEVICE_TABLE(acenic_pci_tbl
) = {
137 { PCI_VENDOR_ID_ALTEON
, PCI_DEVICE_ID_ALTEON_ACENIC_FIBRE
,
138 PCI_ANY_ID
, PCI_ANY_ID
, PCI_CLASS_NETWORK_ETHERNET
<< 8, 0xffff00, },
139 { PCI_VENDOR_ID_ALTEON
, PCI_DEVICE_ID_ALTEON_ACENIC_COPPER
,
140 PCI_ANY_ID
, PCI_ANY_ID
, PCI_CLASS_NETWORK_ETHERNET
<< 8, 0xffff00, },
141 { PCI_VENDOR_ID_3COM
, PCI_DEVICE_ID_3COM_3C985
,
142 PCI_ANY_ID
, PCI_ANY_ID
, PCI_CLASS_NETWORK_ETHERNET
<< 8, 0xffff00, },
143 { PCI_VENDOR_ID_NETGEAR
, PCI_DEVICE_ID_NETGEAR_GA620
,
144 PCI_ANY_ID
, PCI_ANY_ID
, PCI_CLASS_NETWORK_ETHERNET
<< 8, 0xffff00, },
145 { PCI_VENDOR_ID_NETGEAR
, PCI_DEVICE_ID_NETGEAR_GA620T
,
146 PCI_ANY_ID
, PCI_ANY_ID
, PCI_CLASS_NETWORK_ETHERNET
<< 8, 0xffff00, },
148 * Farallon used the DEC vendor ID on their cards incorrectly,
149 * then later Alteon's ID.
151 { PCI_VENDOR_ID_DEC
, PCI_DEVICE_ID_FARALLON_PN9000SX
,
152 PCI_ANY_ID
, PCI_ANY_ID
, PCI_CLASS_NETWORK_ETHERNET
<< 8, 0xffff00, },
153 { PCI_VENDOR_ID_ALTEON
, PCI_DEVICE_ID_FARALLON_PN9100T
,
154 PCI_ANY_ID
, PCI_ANY_ID
, PCI_CLASS_NETWORK_ETHERNET
<< 8, 0xffff00, },
155 { PCI_VENDOR_ID_SGI
, PCI_DEVICE_ID_SGI_ACENIC
,
156 PCI_ANY_ID
, PCI_ANY_ID
, PCI_CLASS_NETWORK_ETHERNET
<< 8, 0xffff00, },
159 MODULE_DEVICE_TABLE(pci
, acenic_pci_tbl
);
161 #define ace_sync_irq(irq) synchronize_irq(irq)
163 #ifndef offset_in_page
164 #define offset_in_page(ptr) ((unsigned long)(ptr) & ~PAGE_MASK)
167 #define ACE_MAX_MOD_PARMS 8
168 #define BOARD_IDX_STATIC 0
169 #define BOARD_IDX_OVERFLOW -1
174 * These must be defined before the firmware is included.
176 #define MAX_TEXT_LEN 96*1024
177 #define MAX_RODATA_LEN 8*1024
178 #define MAX_DATA_LEN 2*1024
180 #ifndef tigon2FwReleaseLocal
181 #define tigon2FwReleaseLocal 0
185 * This driver currently supports Tigon I and Tigon II based cards
186 * including the Alteon AceNIC, the 3Com 3C985[B] and NetGear
187 * GA620. The driver should also work on the SGI, DEC and Farallon
188 * versions of the card, however I have not been able to test that
191 * This card is really neat, it supports receive hardware checksumming
192 * and jumbo frames (up to 9000 bytes) and does a lot of work in the
193 * firmware. Also the programming interface is quite neat, except for
194 * the parts dealing with the i2c eeprom on the card ;-)
196 * Using jumbo frames:
198 * To enable jumbo frames, simply specify an mtu between 1500 and 9000
199 * bytes to ifconfig. Jumbo frames can be enabled or disabled at any time
200 * by running `ifconfig eth<X> mtu <MTU>' with <X> being the Ethernet
201 * interface number and <MTU> being the MTU value.
205 * When compiled as a loadable module, the driver allows for a number
206 * of module parameters to be specified. The driver supports the
207 * following module parameters:
209 * trace=<val> - Firmware trace level. This requires special traced
210 * firmware to replace the firmware supplied with
211 * the driver - for debugging purposes only.
213 * link=<val> - Link state. Normally you want to use the default link
214 * parameters set by the driver. This can be used to
215 * override these in case your switch doesn't negotiate
216 * the link properly. Valid values are:
217 * 0x0001 - Force half duplex link.
218 * 0x0002 - Do not negotiate line speed with the other end.
219 * 0x0010 - 10Mbit/sec link.
220 * 0x0020 - 100Mbit/sec link.
221 * 0x0040 - 1000Mbit/sec link.
222 * 0x0100 - Do not negotiate flow control.
223 * 0x0200 - Enable RX flow control Y
224 * 0x0400 - Enable TX flow control Y (Tigon II NICs only).
225 * Default value is 0x0270, ie. enable link+flow
226 * control negotiation. Negotiating the highest
227 * possible link speed with RX flow control enabled.
229 * When disabling link speed negotiation, only one link
230 * speed is allowed to be specified!
232 * tx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
233 * to wait for more packets to arive before
234 * interrupting the host, from the time the first
237 * rx_coal_tick=<val> - number of coalescing clock ticks (us) allowed
238 * to wait for more packets to arive in the transmit ring,
239 * before interrupting the host, after transmitting the
240 * first packet in the ring.
242 * max_tx_desc=<val> - maximum number of transmit descriptors
243 * (packets) transmitted before interrupting the host.
245 * max_rx_desc=<val> - maximum number of receive descriptors
246 * (packets) received before interrupting the host.
248 * tx_ratio=<val> - 7 bit value (0 - 63) specifying the split in 64th
249 * increments of the NIC's on board memory to be used for
250 * transmit and receive buffers. For the 1MB NIC app. 800KB
251 * is available, on the 1/2MB NIC app. 300KB is available.
252 * 68KB will always be available as a minimum for both
253 * directions. The default value is a 50/50 split.
254 * dis_pci_mem_inval=<val> - disable PCI memory write and invalidate
255 * operations, default (1) is to always disable this as
256 * that is what Alteon does on NT. I have not been able
257 * to measure any real performance differences with
258 * this on my systems. Set <val>=0 if you want to
259 * enable these operations.
261 * If you use more than one NIC, specify the parameters for the
262 * individual NICs with a comma, ie. trace=0,0x00001fff,0 you want to
263 * run tracing on NIC #2 but not on NIC #1 and #3.
267 * - Proper multicast support.
268 * - NIC dump support.
269 * - More tuning parameters.
271 * The mini ring is not used under Linux and I am not sure it makes sense
272 * to actually use it.
274 * New interrupt handler strategy:
276 * The old interrupt handler worked using the traditional method of
277 * replacing an skbuff with a new one when a packet arrives. However
278 * the rx rings do not need to contain a static number of buffer
279 * descriptors, thus it makes sense to move the memory allocation out
280 * of the main interrupt handler and do it in a bottom half handler
281 * and only allocate new buffers when the number of buffers in the
282 * ring is below a certain threshold. In order to avoid starving the
283 * NIC under heavy load it is however necessary to force allocation
284 * when hitting a minimum threshold. The strategy for alloction is as
287 * RX_LOW_BUF_THRES - allocate buffers in the bottom half
288 * RX_PANIC_LOW_THRES - we are very low on buffers, allocate
289 * the buffers in the interrupt handler
290 * RX_RING_THRES - maximum number of buffers in the rx ring
291 * RX_MINI_THRES - maximum number of buffers in the mini ring
292 * RX_JUMBO_THRES - maximum number of buffers in the jumbo ring
294 * One advantagous side effect of this allocation approach is that the
295 * entire rx processing can be done without holding any spin lock
296 * since the rx rings and registers are totally independent of the tx
297 * ring and its registers. This of course includes the kmalloc's of
298 * new skb's. Thus start_xmit can run in parallel with rx processing
299 * and the memory allocation on SMP systems.
301 * Note that running the skb reallocation in a bottom half opens up
302 * another can of races which needs to be handled properly. In
303 * particular it can happen that the interrupt handler tries to run
304 * the reallocation while the bottom half is either running on another
305 * CPU or was interrupted on the same CPU. To get around this the
306 * driver uses bitops to prevent the reallocation routines from being
309 * TX handling can also be done without holding any spin lock, wheee
310 * this is fun! since tx_ret_csm is only written to by the interrupt
311 * handler. The case to be aware of is when shutting down the device
312 * and cleaning up where it is necessary to make sure that
313 * start_xmit() is not running while this is happening. Well DaveM
314 * informs me that this case is already protected against ... bye bye
315 * Mr. Spin Lock, it was nice to know you.
317 * TX interrupts are now partly disabled so the NIC will only generate
318 * TX interrupts for the number of coal ticks, not for the number of
319 * TX packets in the queue. This should reduce the number of TX only,
320 * ie. when no RX processing is done, interrupts seen.
324 * Threshold values for RX buffer allocation - the low water marks for
325 * when to start refilling the rings are set to 75% of the ring
326 * sizes. It seems to make sense to refill the rings entirely from the
327 * intrrupt handler once it gets below the panic threshold, that way
328 * we don't risk that the refilling is moved to another CPU when the
329 * one running the interrupt handler just got the slab code hot in its
332 #define RX_RING_SIZE 72
333 #define RX_MINI_SIZE 64
334 #define RX_JUMBO_SIZE 48
336 #define RX_PANIC_STD_THRES 16
337 #define RX_PANIC_STD_REFILL (3*RX_PANIC_STD_THRES)/2
338 #define RX_LOW_STD_THRES (3*RX_RING_SIZE)/4
339 #define RX_PANIC_MINI_THRES 12
340 #define RX_PANIC_MINI_REFILL (3*RX_PANIC_MINI_THRES)/2
341 #define RX_LOW_MINI_THRES (3*RX_MINI_SIZE)/4
342 #define RX_PANIC_JUMBO_THRES 6
343 #define RX_PANIC_JUMBO_REFILL (3*RX_PANIC_JUMBO_THRES)/2
344 #define RX_LOW_JUMBO_THRES (3*RX_JUMBO_SIZE)/4
348 * Size of the mini ring entries, basically these just should be big
349 * enough to take TCP ACKs
351 #define ACE_MINI_SIZE 100
353 #define ACE_MINI_BUFSIZE ACE_MINI_SIZE
354 #define ACE_STD_BUFSIZE (ACE_STD_MTU + ETH_HLEN + 4)
355 #define ACE_JUMBO_BUFSIZE (ACE_JUMBO_MTU + ETH_HLEN + 4)
358 * There seems to be a magic difference in the effect between 995 and 996
359 * but little difference between 900 and 995 ... no idea why.
361 * There is now a default set of tuning parameters which is set, depending
362 * on whether or not the user enables Jumbo frames. It's assumed that if
363 * Jumbo frames are enabled, the user wants optimal tuning for that case.
365 #define DEF_TX_COAL 400 /* 996 */
366 #define DEF_TX_MAX_DESC 60 /* was 40 */
367 #define DEF_RX_COAL 120 /* 1000 */
368 #define DEF_RX_MAX_DESC 25
369 #define DEF_TX_RATIO 21 /* 24 */
371 #define DEF_JUMBO_TX_COAL 20
372 #define DEF_JUMBO_TX_MAX_DESC 60
373 #define DEF_JUMBO_RX_COAL 30
374 #define DEF_JUMBO_RX_MAX_DESC 6
375 #define DEF_JUMBO_TX_RATIO 21
377 #if tigon2FwReleaseLocal < 20001118
379 * Standard firmware and early modifications duplicate
380 * IRQ load without this flag (coal timer is never reset).
381 * Note that with this flag tx_coal should be less than
382 * time to xmit full tx ring.
383 * 400usec is not so bad for tx ring size of 128.
385 #define TX_COAL_INTS_ONLY 1 /* worth it */
388 * With modified firmware, this is not necessary, but still useful.
390 #define TX_COAL_INTS_ONLY 1
394 #define DEF_STAT (2 * TICKS_PER_SEC)
397 static int link_state
[ACE_MAX_MOD_PARMS
];
398 static int trace
[ACE_MAX_MOD_PARMS
];
399 static int tx_coal_tick
[ACE_MAX_MOD_PARMS
];
400 static int rx_coal_tick
[ACE_MAX_MOD_PARMS
];
401 static int max_tx_desc
[ACE_MAX_MOD_PARMS
];
402 static int max_rx_desc
[ACE_MAX_MOD_PARMS
];
403 static int tx_ratio
[ACE_MAX_MOD_PARMS
];
404 static int dis_pci_mem_inval
[ACE_MAX_MOD_PARMS
] = {1, 1, 1, 1, 1, 1, 1, 1};
406 MODULE_AUTHOR("Jes Sorensen <jes@trained-monkey.org>");
407 MODULE_LICENSE("GPL");
408 MODULE_DESCRIPTION("AceNIC/3C985/GA620 Gigabit Ethernet driver");
409 #ifndef CONFIG_ACENIC_OMIT_TIGON_I
410 MODULE_FIRMWARE("acenic/tg1.bin");
412 MODULE_FIRMWARE("acenic/tg2.bin");
414 module_param_array_named(link
, link_state
, int, NULL
, 0);
415 module_param_array(trace
, int, NULL
, 0);
416 module_param_array(tx_coal_tick
, int, NULL
, 0);
417 module_param_array(max_tx_desc
, int, NULL
, 0);
418 module_param_array(rx_coal_tick
, int, NULL
, 0);
419 module_param_array(max_rx_desc
, int, NULL
, 0);
420 module_param_array(tx_ratio
, int, NULL
, 0);
421 MODULE_PARM_DESC(link
, "AceNIC/3C985/NetGear link state");
422 MODULE_PARM_DESC(trace
, "AceNIC/3C985/NetGear firmware trace level");
423 MODULE_PARM_DESC(tx_coal_tick
, "AceNIC/3C985/GA620 max clock ticks to wait from first tx descriptor arrives");
424 MODULE_PARM_DESC(max_tx_desc
, "AceNIC/3C985/GA620 max number of transmit descriptors to wait");
425 MODULE_PARM_DESC(rx_coal_tick
, "AceNIC/3C985/GA620 max clock ticks to wait from first rx descriptor arrives");
426 MODULE_PARM_DESC(max_rx_desc
, "AceNIC/3C985/GA620 max number of receive descriptors to wait");
427 MODULE_PARM_DESC(tx_ratio
, "AceNIC/3C985/GA620 ratio of NIC memory used for TX/RX descriptors (range 0-63)");
430 static const char version
[] __devinitconst
=
431 "acenic.c: v0.92 08/05/2002 Jes Sorensen, linux-acenic@SunSITE.dk\n"
432 " http://home.cern.ch/~jes/gige/acenic.html\n";
434 static int ace_get_settings(struct net_device
*, struct ethtool_cmd
*);
435 static int ace_set_settings(struct net_device
*, struct ethtool_cmd
*);
436 static void ace_get_drvinfo(struct net_device
*, struct ethtool_drvinfo
*);
438 static const struct ethtool_ops ace_ethtool_ops
= {
439 .get_settings
= ace_get_settings
,
440 .set_settings
= ace_set_settings
,
441 .get_drvinfo
= ace_get_drvinfo
,
444 static void ace_watchdog(struct net_device
*dev
);
446 static const struct net_device_ops ace_netdev_ops
= {
447 .ndo_open
= ace_open
,
448 .ndo_stop
= ace_close
,
449 .ndo_tx_timeout
= ace_watchdog
,
450 .ndo_get_stats
= ace_get_stats
,
451 .ndo_start_xmit
= ace_start_xmit
,
452 .ndo_set_rx_mode
= ace_set_multicast_list
,
453 .ndo_validate_addr
= eth_validate_addr
,
454 .ndo_set_mac_address
= ace_set_mac_addr
,
455 .ndo_change_mtu
= ace_change_mtu
,
458 static int __devinit
acenic_probe_one(struct pci_dev
*pdev
,
459 const struct pci_device_id
*id
)
461 struct net_device
*dev
;
462 struct ace_private
*ap
;
463 static int boards_found
;
465 dev
= alloc_etherdev(sizeof(struct ace_private
));
467 printk(KERN_ERR
"acenic: Unable to allocate "
468 "net_device structure!\n");
472 SET_NETDEV_DEV(dev
, &pdev
->dev
);
474 ap
= netdev_priv(dev
);
476 ap
->name
= pci_name(pdev
);
478 dev
->features
|= NETIF_F_SG
| NETIF_F_IP_CSUM
;
479 dev
->features
|= NETIF_F_HW_VLAN_TX
| NETIF_F_HW_VLAN_RX
;
481 dev
->watchdog_timeo
= 5*HZ
;
483 dev
->netdev_ops
= &ace_netdev_ops
;
484 SET_ETHTOOL_OPS(dev
, &ace_ethtool_ops
);
486 /* we only display this string ONCE */
490 if (pci_enable_device(pdev
))
491 goto fail_free_netdev
;
494 * Enable master mode before we start playing with the
495 * pci_command word since pci_set_master() will modify
498 pci_set_master(pdev
);
500 pci_read_config_word(pdev
, PCI_COMMAND
, &ap
->pci_command
);
502 /* OpenFirmware on Mac's does not set this - DOH.. */
503 if (!(ap
->pci_command
& PCI_COMMAND_MEMORY
)) {
504 printk(KERN_INFO
"%s: Enabling PCI Memory Mapped "
505 "access - was not enabled by BIOS/Firmware\n",
507 ap
->pci_command
= ap
->pci_command
| PCI_COMMAND_MEMORY
;
508 pci_write_config_word(ap
->pdev
, PCI_COMMAND
,
513 pci_read_config_byte(pdev
, PCI_LATENCY_TIMER
, &ap
->pci_latency
);
514 if (ap
->pci_latency
<= 0x40) {
515 ap
->pci_latency
= 0x40;
516 pci_write_config_byte(pdev
, PCI_LATENCY_TIMER
, ap
->pci_latency
);
520 * Remap the regs into kernel space - this is abuse of
521 * dev->base_addr since it was means for I/O port
522 * addresses but who gives a damn.
524 dev
->base_addr
= pci_resource_start(pdev
, 0);
525 ap
->regs
= ioremap(dev
->base_addr
, 0x4000);
527 printk(KERN_ERR
"%s: Unable to map I/O register, "
528 "AceNIC %i will be disabled.\n",
529 ap
->name
, boards_found
);
530 goto fail_free_netdev
;
533 switch(pdev
->vendor
) {
534 case PCI_VENDOR_ID_ALTEON
:
535 if (pdev
->device
== PCI_DEVICE_ID_FARALLON_PN9100T
) {
536 printk(KERN_INFO
"%s: Farallon PN9100-T ",
539 printk(KERN_INFO
"%s: Alteon AceNIC ",
543 case PCI_VENDOR_ID_3COM
:
544 printk(KERN_INFO
"%s: 3Com 3C985 ", ap
->name
);
546 case PCI_VENDOR_ID_NETGEAR
:
547 printk(KERN_INFO
"%s: NetGear GA620 ", ap
->name
);
549 case PCI_VENDOR_ID_DEC
:
550 if (pdev
->device
== PCI_DEVICE_ID_FARALLON_PN9000SX
) {
551 printk(KERN_INFO
"%s: Farallon PN9000-SX ",
555 case PCI_VENDOR_ID_SGI
:
556 printk(KERN_INFO
"%s: SGI AceNIC ", ap
->name
);
559 printk(KERN_INFO
"%s: Unknown AceNIC ", ap
->name
);
563 printk("Gigabit Ethernet at 0x%08lx, ", dev
->base_addr
);
564 printk("irq %d\n", pdev
->irq
);
566 #ifdef CONFIG_ACENIC_OMIT_TIGON_I
567 if ((readl(&ap
->regs
->HostCtrl
) >> 28) == 4) {
568 printk(KERN_ERR
"%s: Driver compiled without Tigon I"
569 " support - NIC disabled\n", dev
->name
);
574 if (ace_allocate_descriptors(dev
))
575 goto fail_free_netdev
;
578 if (boards_found
>= ACE_MAX_MOD_PARMS
)
579 ap
->board_idx
= BOARD_IDX_OVERFLOW
;
581 ap
->board_idx
= boards_found
;
583 ap
->board_idx
= BOARD_IDX_STATIC
;
587 goto fail_free_netdev
;
589 if (register_netdev(dev
)) {
590 printk(KERN_ERR
"acenic: device registration failed\n");
593 ap
->name
= dev
->name
;
595 if (ap
->pci_using_dac
)
596 dev
->features
|= NETIF_F_HIGHDMA
;
598 pci_set_drvdata(pdev
, dev
);
604 ace_init_cleanup(dev
);
610 static void __devexit
acenic_remove_one(struct pci_dev
*pdev
)
612 struct net_device
*dev
= pci_get_drvdata(pdev
);
613 struct ace_private
*ap
= netdev_priv(dev
);
614 struct ace_regs __iomem
*regs
= ap
->regs
;
617 unregister_netdev(dev
);
619 writel(readl(®s
->CpuCtrl
) | CPU_HALT
, ®s
->CpuCtrl
);
620 if (ap
->version
>= 2)
621 writel(readl(®s
->CpuBCtrl
) | CPU_HALT
, ®s
->CpuBCtrl
);
624 * This clears any pending interrupts
626 writel(1, ®s
->Mb0Lo
);
627 readl(®s
->CpuCtrl
); /* flush */
630 * Make sure no other CPUs are processing interrupts
631 * on the card before the buffers are being released.
632 * Otherwise one might experience some `interesting'
635 * Then release the RX buffers - jumbo buffers were
636 * already released in ace_close().
638 ace_sync_irq(dev
->irq
);
640 for (i
= 0; i
< RX_STD_RING_ENTRIES
; i
++) {
641 struct sk_buff
*skb
= ap
->skb
->rx_std_skbuff
[i
].skb
;
644 struct ring_info
*ringp
;
647 ringp
= &ap
->skb
->rx_std_skbuff
[i
];
648 mapping
= dma_unmap_addr(ringp
, mapping
);
649 pci_unmap_page(ap
->pdev
, mapping
,
653 ap
->rx_std_ring
[i
].size
= 0;
654 ap
->skb
->rx_std_skbuff
[i
].skb
= NULL
;
659 if (ap
->version
>= 2) {
660 for (i
= 0; i
< RX_MINI_RING_ENTRIES
; i
++) {
661 struct sk_buff
*skb
= ap
->skb
->rx_mini_skbuff
[i
].skb
;
664 struct ring_info
*ringp
;
667 ringp
= &ap
->skb
->rx_mini_skbuff
[i
];
668 mapping
= dma_unmap_addr(ringp
,mapping
);
669 pci_unmap_page(ap
->pdev
, mapping
,
673 ap
->rx_mini_ring
[i
].size
= 0;
674 ap
->skb
->rx_mini_skbuff
[i
].skb
= NULL
;
680 for (i
= 0; i
< RX_JUMBO_RING_ENTRIES
; i
++) {
681 struct sk_buff
*skb
= ap
->skb
->rx_jumbo_skbuff
[i
].skb
;
683 struct ring_info
*ringp
;
686 ringp
= &ap
->skb
->rx_jumbo_skbuff
[i
];
687 mapping
= dma_unmap_addr(ringp
, mapping
);
688 pci_unmap_page(ap
->pdev
, mapping
,
692 ap
->rx_jumbo_ring
[i
].size
= 0;
693 ap
->skb
->rx_jumbo_skbuff
[i
].skb
= NULL
;
698 ace_init_cleanup(dev
);
702 static struct pci_driver acenic_pci_driver
= {
704 .id_table
= acenic_pci_tbl
,
705 .probe
= acenic_probe_one
,
706 .remove
= __devexit_p(acenic_remove_one
),
709 static int __init
acenic_init(void)
711 return pci_register_driver(&acenic_pci_driver
);
714 static void __exit
acenic_exit(void)
716 pci_unregister_driver(&acenic_pci_driver
);
719 module_init(acenic_init
);
720 module_exit(acenic_exit
);
722 static void ace_free_descriptors(struct net_device
*dev
)
724 struct ace_private
*ap
= netdev_priv(dev
);
727 if (ap
->rx_std_ring
!= NULL
) {
728 size
= (sizeof(struct rx_desc
) *
729 (RX_STD_RING_ENTRIES
+
730 RX_JUMBO_RING_ENTRIES
+
731 RX_MINI_RING_ENTRIES
+
732 RX_RETURN_RING_ENTRIES
));
733 pci_free_consistent(ap
->pdev
, size
, ap
->rx_std_ring
,
734 ap
->rx_ring_base_dma
);
735 ap
->rx_std_ring
= NULL
;
736 ap
->rx_jumbo_ring
= NULL
;
737 ap
->rx_mini_ring
= NULL
;
738 ap
->rx_return_ring
= NULL
;
740 if (ap
->evt_ring
!= NULL
) {
741 size
= (sizeof(struct event
) * EVT_RING_ENTRIES
);
742 pci_free_consistent(ap
->pdev
, size
, ap
->evt_ring
,
746 if (ap
->tx_ring
!= NULL
&& !ACE_IS_TIGON_I(ap
)) {
747 size
= (sizeof(struct tx_desc
) * MAX_TX_RING_ENTRIES
);
748 pci_free_consistent(ap
->pdev
, size
, ap
->tx_ring
,
753 if (ap
->evt_prd
!= NULL
) {
754 pci_free_consistent(ap
->pdev
, sizeof(u32
),
755 (void *)ap
->evt_prd
, ap
->evt_prd_dma
);
758 if (ap
->rx_ret_prd
!= NULL
) {
759 pci_free_consistent(ap
->pdev
, sizeof(u32
),
760 (void *)ap
->rx_ret_prd
,
762 ap
->rx_ret_prd
= NULL
;
764 if (ap
->tx_csm
!= NULL
) {
765 pci_free_consistent(ap
->pdev
, sizeof(u32
),
766 (void *)ap
->tx_csm
, ap
->tx_csm_dma
);
772 static int ace_allocate_descriptors(struct net_device
*dev
)
774 struct ace_private
*ap
= netdev_priv(dev
);
777 size
= (sizeof(struct rx_desc
) *
778 (RX_STD_RING_ENTRIES
+
779 RX_JUMBO_RING_ENTRIES
+
780 RX_MINI_RING_ENTRIES
+
781 RX_RETURN_RING_ENTRIES
));
783 ap
->rx_std_ring
= pci_alloc_consistent(ap
->pdev
, size
,
784 &ap
->rx_ring_base_dma
);
785 if (ap
->rx_std_ring
== NULL
)
788 ap
->rx_jumbo_ring
= ap
->rx_std_ring
+ RX_STD_RING_ENTRIES
;
789 ap
->rx_mini_ring
= ap
->rx_jumbo_ring
+ RX_JUMBO_RING_ENTRIES
;
790 ap
->rx_return_ring
= ap
->rx_mini_ring
+ RX_MINI_RING_ENTRIES
;
792 size
= (sizeof(struct event
) * EVT_RING_ENTRIES
);
794 ap
->evt_ring
= pci_alloc_consistent(ap
->pdev
, size
, &ap
->evt_ring_dma
);
796 if (ap
->evt_ring
== NULL
)
800 * Only allocate a host TX ring for the Tigon II, the Tigon I
801 * has to use PCI registers for this ;-(
803 if (!ACE_IS_TIGON_I(ap
)) {
804 size
= (sizeof(struct tx_desc
) * MAX_TX_RING_ENTRIES
);
806 ap
->tx_ring
= pci_alloc_consistent(ap
->pdev
, size
,
809 if (ap
->tx_ring
== NULL
)
813 ap
->evt_prd
= pci_alloc_consistent(ap
->pdev
, sizeof(u32
),
815 if (ap
->evt_prd
== NULL
)
818 ap
->rx_ret_prd
= pci_alloc_consistent(ap
->pdev
, sizeof(u32
),
819 &ap
->rx_ret_prd_dma
);
820 if (ap
->rx_ret_prd
== NULL
)
823 ap
->tx_csm
= pci_alloc_consistent(ap
->pdev
, sizeof(u32
),
825 if (ap
->tx_csm
== NULL
)
832 ace_init_cleanup(dev
);
838 * Generic cleanup handling data allocated during init. Used when the
839 * module is unloaded or if an error occurs during initialization
841 static void ace_init_cleanup(struct net_device
*dev
)
843 struct ace_private
*ap
;
845 ap
= netdev_priv(dev
);
847 ace_free_descriptors(dev
);
850 pci_free_consistent(ap
->pdev
, sizeof(struct ace_info
),
851 ap
->info
, ap
->info_dma
);
853 kfree(ap
->trace_buf
);
856 free_irq(dev
->irq
, dev
);
863 * Commands are considered to be slow.
865 static inline void ace_issue_cmd(struct ace_regs __iomem
*regs
, struct cmd
*cmd
)
869 idx
= readl(®s
->CmdPrd
);
871 writel(*(u32
*)(cmd
), ®s
->CmdRng
[idx
]);
872 idx
= (idx
+ 1) % CMD_RING_ENTRIES
;
874 writel(idx
, ®s
->CmdPrd
);
878 static int __devinit
ace_init(struct net_device
*dev
)
880 struct ace_private
*ap
;
881 struct ace_regs __iomem
*regs
;
882 struct ace_info
*info
= NULL
;
883 struct pci_dev
*pdev
;
886 u32 tig_ver
, mac1
, mac2
, tmp
, pci_state
;
887 int board_idx
, ecode
= 0;
889 unsigned char cache_size
;
891 ap
= netdev_priv(dev
);
894 board_idx
= ap
->board_idx
;
897 * aman@sgi.com - its useful to do a NIC reset here to
898 * address the `Firmware not running' problem subsequent
899 * to any crashes involving the NIC
901 writel(HW_RESET
| (HW_RESET
<< 24), ®s
->HostCtrl
);
902 readl(®s
->HostCtrl
); /* PCI write posting */
906 * Don't access any other registers before this point!
910 * This will most likely need BYTE_SWAP once we switch
911 * to using __raw_writel()
913 writel((WORD_SWAP
| CLR_INT
| ((WORD_SWAP
| CLR_INT
) << 24)),
916 writel((CLR_INT
| WORD_SWAP
| ((CLR_INT
| WORD_SWAP
) << 24)),
919 readl(®s
->HostCtrl
); /* PCI write posting */
922 * Stop the NIC CPU and clear pending interrupts
924 writel(readl(®s
->CpuCtrl
) | CPU_HALT
, ®s
->CpuCtrl
);
925 readl(®s
->CpuCtrl
); /* PCI write posting */
926 writel(0, ®s
->Mb0Lo
);
928 tig_ver
= readl(®s
->HostCtrl
) >> 28;
931 #ifndef CONFIG_ACENIC_OMIT_TIGON_I
934 printk(KERN_INFO
" Tigon I (Rev. %i), Firmware: %i.%i.%i, ",
935 tig_ver
, ap
->firmware_major
, ap
->firmware_minor
,
937 writel(0, ®s
->LocalCtrl
);
939 ap
->tx_ring_entries
= TIGON_I_TX_RING_ENTRIES
;
943 printk(KERN_INFO
" Tigon II (Rev. %i), Firmware: %i.%i.%i, ",
944 tig_ver
, ap
->firmware_major
, ap
->firmware_minor
,
946 writel(readl(®s
->CpuBCtrl
) | CPU_HALT
, ®s
->CpuBCtrl
);
947 readl(®s
->CpuBCtrl
); /* PCI write posting */
949 * The SRAM bank size does _not_ indicate the amount
950 * of memory on the card, it controls the _bank_ size!
951 * Ie. a 1MB AceNIC will have two banks of 512KB.
953 writel(SRAM_BANK_512K
, ®s
->LocalCtrl
);
954 writel(SYNC_SRAM_TIMING
, ®s
->MiscCfg
);
956 ap
->tx_ring_entries
= MAX_TX_RING_ENTRIES
;
959 printk(KERN_WARNING
" Unsupported Tigon version detected "
966 * ModeStat _must_ be set after the SRAM settings as this change
967 * seems to corrupt the ModeStat and possible other registers.
968 * The SRAM settings survive resets and setting it to the same
969 * value a second time works as well. This is what caused the
970 * `Firmware not running' problem on the Tigon II.
973 writel(ACE_BYTE_SWAP_DMA
| ACE_WARN
| ACE_FATAL
| ACE_BYTE_SWAP_BD
|
974 ACE_WORD_SWAP_BD
| ACE_NO_JUMBO_FRAG
, ®s
->ModeStat
);
976 writel(ACE_BYTE_SWAP_DMA
| ACE_WARN
| ACE_FATAL
|
977 ACE_WORD_SWAP_BD
| ACE_NO_JUMBO_FRAG
, ®s
->ModeStat
);
979 readl(®s
->ModeStat
); /* PCI write posting */
982 for(i
= 0; i
< 4; i
++) {
986 t
= read_eeprom_byte(dev
, 0x8c+i
);
994 for(i
= 4; i
< 8; i
++) {
998 t
= read_eeprom_byte(dev
, 0x8c+i
);
1006 writel(mac1
, ®s
->MacAddrHi
);
1007 writel(mac2
, ®s
->MacAddrLo
);
1009 dev
->dev_addr
[0] = (mac1
>> 8) & 0xff;
1010 dev
->dev_addr
[1] = mac1
& 0xff;
1011 dev
->dev_addr
[2] = (mac2
>> 24) & 0xff;
1012 dev
->dev_addr
[3] = (mac2
>> 16) & 0xff;
1013 dev
->dev_addr
[4] = (mac2
>> 8) & 0xff;
1014 dev
->dev_addr
[5] = mac2
& 0xff;
1016 printk("MAC: %pM\n", dev
->dev_addr
);
1019 * Looks like this is necessary to deal with on all architectures,
1020 * even this %$#%$# N440BX Intel based thing doesn't get it right.
1021 * Ie. having two NICs in the machine, one will have the cache
1022 * line set at boot time, the other will not.
1025 pci_read_config_byte(pdev
, PCI_CACHE_LINE_SIZE
, &cache_size
);
1027 if (cache_size
!= SMP_CACHE_BYTES
) {
1028 printk(KERN_INFO
" PCI cache line size set incorrectly "
1029 "(%i bytes) by BIOS/FW, ", cache_size
);
1030 if (cache_size
> SMP_CACHE_BYTES
)
1031 printk("expecting %i\n", SMP_CACHE_BYTES
);
1033 printk("correcting to %i\n", SMP_CACHE_BYTES
);
1034 pci_write_config_byte(pdev
, PCI_CACHE_LINE_SIZE
,
1035 SMP_CACHE_BYTES
>> 2);
1039 pci_state
= readl(®s
->PciState
);
1040 printk(KERN_INFO
" PCI bus width: %i bits, speed: %iMHz, "
1041 "latency: %i clks\n",
1042 (pci_state
& PCI_32BIT
) ? 32 : 64,
1043 (pci_state
& PCI_66MHZ
) ? 66 : 33,
1047 * Set the max DMA transfer size. Seems that for most systems
1048 * the performance is better when no MAX parameter is
1049 * set. However for systems enabling PCI write and invalidate,
1050 * DMA writes must be set to the L1 cache line size to get
1051 * optimal performance.
1053 * The default is now to turn the PCI write and invalidate off
1054 * - that is what Alteon does for NT.
1056 tmp
= READ_CMD_MEM
| WRITE_CMD_MEM
;
1057 if (ap
->version
>= 2) {
1058 tmp
|= (MEM_READ_MULTIPLE
| (pci_state
& PCI_66MHZ
));
1060 * Tuning parameters only supported for 8 cards
1062 if (board_idx
== BOARD_IDX_OVERFLOW
||
1063 dis_pci_mem_inval
[board_idx
]) {
1064 if (ap
->pci_command
& PCI_COMMAND_INVALIDATE
) {
1065 ap
->pci_command
&= ~PCI_COMMAND_INVALIDATE
;
1066 pci_write_config_word(pdev
, PCI_COMMAND
,
1068 printk(KERN_INFO
" Disabling PCI memory "
1069 "write and invalidate\n");
1071 } else if (ap
->pci_command
& PCI_COMMAND_INVALIDATE
) {
1072 printk(KERN_INFO
" PCI memory write & invalidate "
1073 "enabled by BIOS, enabling counter measures\n");
1075 switch(SMP_CACHE_BYTES
) {
1077 tmp
|= DMA_WRITE_MAX_16
;
1080 tmp
|= DMA_WRITE_MAX_32
;
1083 tmp
|= DMA_WRITE_MAX_64
;
1086 tmp
|= DMA_WRITE_MAX_128
;
1089 printk(KERN_INFO
" Cache line size %i not "
1090 "supported, PCI write and invalidate "
1091 "disabled\n", SMP_CACHE_BYTES
);
1092 ap
->pci_command
&= ~PCI_COMMAND_INVALIDATE
;
1093 pci_write_config_word(pdev
, PCI_COMMAND
,
1101 * On this platform, we know what the best dma settings
1102 * are. We use 64-byte maximum bursts, because if we
1103 * burst larger than the cache line size (or even cross
1104 * a 64byte boundary in a single burst) the UltraSparc
1105 * PCI controller will disconnect at 64-byte multiples.
1107 * Read-multiple will be properly enabled above, and when
1108 * set will give the PCI controller proper hints about
1111 tmp
&= ~DMA_READ_WRITE_MASK
;
1112 tmp
|= DMA_READ_MAX_64
;
1113 tmp
|= DMA_WRITE_MAX_64
;
1116 tmp
&= ~DMA_READ_WRITE_MASK
;
1117 tmp
|= DMA_READ_MAX_128
;
1119 * All the docs say MUST NOT. Well, I did.
1120 * Nothing terrible happens, if we load wrong size.
1121 * Bit w&i still works better!
1123 tmp
|= DMA_WRITE_MAX_128
;
1125 writel(tmp
, ®s
->PciState
);
1129 * The Host PCI bus controller driver has to set FBB.
1130 * If all devices on that PCI bus support FBB, then the controller
1131 * can enable FBB support in the Host PCI Bus controller (or on
1132 * the PCI-PCI bridge if that applies).
1136 * I have received reports from people having problems when this
1139 if (!(ap
->pci_command
& PCI_COMMAND_FAST_BACK
)) {
1140 printk(KERN_INFO
" Enabling PCI Fast Back to Back\n");
1141 ap
->pci_command
|= PCI_COMMAND_FAST_BACK
;
1142 pci_write_config_word(pdev
, PCI_COMMAND
, ap
->pci_command
);
1147 * Configure DMA attributes.
1149 if (!pci_set_dma_mask(pdev
, DMA_BIT_MASK(64))) {
1150 ap
->pci_using_dac
= 1;
1151 } else if (!pci_set_dma_mask(pdev
, DMA_BIT_MASK(32))) {
1152 ap
->pci_using_dac
= 0;
1159 * Initialize the generic info block and the command+event rings
1160 * and the control blocks for the transmit and receive rings
1161 * as they need to be setup once and for all.
1163 if (!(info
= pci_alloc_consistent(ap
->pdev
, sizeof(struct ace_info
),
1171 * Get the memory for the skb rings.
1173 if (!(ap
->skb
= kmalloc(sizeof(struct ace_skb
), GFP_KERNEL
))) {
1178 ecode
= request_irq(pdev
->irq
, ace_interrupt
, IRQF_SHARED
,
1181 printk(KERN_WARNING
"%s: Requested IRQ %d is busy\n",
1182 DRV_NAME
, pdev
->irq
);
1185 dev
->irq
= pdev
->irq
;
1188 spin_lock_init(&ap
->debug_lock
);
1189 ap
->last_tx
= ACE_TX_RING_ENTRIES(ap
) - 1;
1190 ap
->last_std_rx
= 0;
1191 ap
->last_mini_rx
= 0;
1194 memset(ap
->info
, 0, sizeof(struct ace_info
));
1195 memset(ap
->skb
, 0, sizeof(struct ace_skb
));
1197 ecode
= ace_load_firmware(dev
);
1203 tmp_ptr
= ap
->info_dma
;
1204 writel(tmp_ptr
>> 32, ®s
->InfoPtrHi
);
1205 writel(tmp_ptr
& 0xffffffff, ®s
->InfoPtrLo
);
1207 memset(ap
->evt_ring
, 0, EVT_RING_ENTRIES
* sizeof(struct event
));
1209 set_aceaddr(&info
->evt_ctrl
.rngptr
, ap
->evt_ring_dma
);
1210 info
->evt_ctrl
.flags
= 0;
1214 set_aceaddr(&info
->evt_prd_ptr
, ap
->evt_prd_dma
);
1215 writel(0, ®s
->EvtCsm
);
1217 set_aceaddr(&info
->cmd_ctrl
.rngptr
, 0x100);
1218 info
->cmd_ctrl
.flags
= 0;
1219 info
->cmd_ctrl
.max_len
= 0;
1221 for (i
= 0; i
< CMD_RING_ENTRIES
; i
++)
1222 writel(0, ®s
->CmdRng
[i
]);
1224 writel(0, ®s
->CmdPrd
);
1225 writel(0, ®s
->CmdCsm
);
1227 tmp_ptr
= ap
->info_dma
;
1228 tmp_ptr
+= (unsigned long) &(((struct ace_info
*)0)->s
.stats
);
1229 set_aceaddr(&info
->stats2_ptr
, (dma_addr_t
) tmp_ptr
);
1231 set_aceaddr(&info
->rx_std_ctrl
.rngptr
, ap
->rx_ring_base_dma
);
1232 info
->rx_std_ctrl
.max_len
= ACE_STD_BUFSIZE
;
1233 info
->rx_std_ctrl
.flags
=
1234 RCB_FLG_TCP_UDP_SUM
| RCB_FLG_NO_PSEUDO_HDR
| RCB_FLG_VLAN_ASSIST
;
1236 memset(ap
->rx_std_ring
, 0,
1237 RX_STD_RING_ENTRIES
* sizeof(struct rx_desc
));
1239 for (i
= 0; i
< RX_STD_RING_ENTRIES
; i
++)
1240 ap
->rx_std_ring
[i
].flags
= BD_FLG_TCP_UDP_SUM
;
1242 ap
->rx_std_skbprd
= 0;
1243 atomic_set(&ap
->cur_rx_bufs
, 0);
1245 set_aceaddr(&info
->rx_jumbo_ctrl
.rngptr
,
1246 (ap
->rx_ring_base_dma
+
1247 (sizeof(struct rx_desc
) * RX_STD_RING_ENTRIES
)));
1248 info
->rx_jumbo_ctrl
.max_len
= 0;
1249 info
->rx_jumbo_ctrl
.flags
=
1250 RCB_FLG_TCP_UDP_SUM
| RCB_FLG_NO_PSEUDO_HDR
| RCB_FLG_VLAN_ASSIST
;
1252 memset(ap
->rx_jumbo_ring
, 0,
1253 RX_JUMBO_RING_ENTRIES
* sizeof(struct rx_desc
));
1255 for (i
= 0; i
< RX_JUMBO_RING_ENTRIES
; i
++)
1256 ap
->rx_jumbo_ring
[i
].flags
= BD_FLG_TCP_UDP_SUM
| BD_FLG_JUMBO
;
1258 ap
->rx_jumbo_skbprd
= 0;
1259 atomic_set(&ap
->cur_jumbo_bufs
, 0);
1261 memset(ap
->rx_mini_ring
, 0,
1262 RX_MINI_RING_ENTRIES
* sizeof(struct rx_desc
));
1264 if (ap
->version
>= 2) {
1265 set_aceaddr(&info
->rx_mini_ctrl
.rngptr
,
1266 (ap
->rx_ring_base_dma
+
1267 (sizeof(struct rx_desc
) *
1268 (RX_STD_RING_ENTRIES
+
1269 RX_JUMBO_RING_ENTRIES
))));
1270 info
->rx_mini_ctrl
.max_len
= ACE_MINI_SIZE
;
1271 info
->rx_mini_ctrl
.flags
=
1272 RCB_FLG_TCP_UDP_SUM
|RCB_FLG_NO_PSEUDO_HDR
|RCB_FLG_VLAN_ASSIST
;
1274 for (i
= 0; i
< RX_MINI_RING_ENTRIES
; i
++)
1275 ap
->rx_mini_ring
[i
].flags
=
1276 BD_FLG_TCP_UDP_SUM
| BD_FLG_MINI
;
1278 set_aceaddr(&info
->rx_mini_ctrl
.rngptr
, 0);
1279 info
->rx_mini_ctrl
.flags
= RCB_FLG_RNG_DISABLE
;
1280 info
->rx_mini_ctrl
.max_len
= 0;
1283 ap
->rx_mini_skbprd
= 0;
1284 atomic_set(&ap
->cur_mini_bufs
, 0);
1286 set_aceaddr(&info
->rx_return_ctrl
.rngptr
,
1287 (ap
->rx_ring_base_dma
+
1288 (sizeof(struct rx_desc
) *
1289 (RX_STD_RING_ENTRIES
+
1290 RX_JUMBO_RING_ENTRIES
+
1291 RX_MINI_RING_ENTRIES
))));
1292 info
->rx_return_ctrl
.flags
= 0;
1293 info
->rx_return_ctrl
.max_len
= RX_RETURN_RING_ENTRIES
;
1295 memset(ap
->rx_return_ring
, 0,
1296 RX_RETURN_RING_ENTRIES
* sizeof(struct rx_desc
));
1298 set_aceaddr(&info
->rx_ret_prd_ptr
, ap
->rx_ret_prd_dma
);
1299 *(ap
->rx_ret_prd
) = 0;
1301 writel(TX_RING_BASE
, ®s
->WinBase
);
1303 if (ACE_IS_TIGON_I(ap
)) {
1304 ap
->tx_ring
= (__force
struct tx_desc
*) regs
->Window
;
1305 for (i
= 0; i
< (TIGON_I_TX_RING_ENTRIES
1306 * sizeof(struct tx_desc
)) / sizeof(u32
); i
++)
1307 writel(0, (__force
void __iomem
*)ap
->tx_ring
+ i
* 4);
1309 set_aceaddr(&info
->tx_ctrl
.rngptr
, TX_RING_BASE
);
1311 memset(ap
->tx_ring
, 0,
1312 MAX_TX_RING_ENTRIES
* sizeof(struct tx_desc
));
1314 set_aceaddr(&info
->tx_ctrl
.rngptr
, ap
->tx_ring_dma
);
1317 info
->tx_ctrl
.max_len
= ACE_TX_RING_ENTRIES(ap
);
1318 tmp
= RCB_FLG_TCP_UDP_SUM
| RCB_FLG_NO_PSEUDO_HDR
| RCB_FLG_VLAN_ASSIST
;
1321 * The Tigon I does not like having the TX ring in host memory ;-(
1323 if (!ACE_IS_TIGON_I(ap
))
1324 tmp
|= RCB_FLG_TX_HOST_RING
;
1325 #if TX_COAL_INTS_ONLY
1326 tmp
|= RCB_FLG_COAL_INT_ONLY
;
1328 info
->tx_ctrl
.flags
= tmp
;
1330 set_aceaddr(&info
->tx_csm_ptr
, ap
->tx_csm_dma
);
1333 * Potential item for tuning parameter
1336 writel(DMA_THRESH_16W
, ®s
->DmaReadCfg
);
1337 writel(DMA_THRESH_16W
, ®s
->DmaWriteCfg
);
1339 writel(DMA_THRESH_8W
, ®s
->DmaReadCfg
);
1340 writel(DMA_THRESH_8W
, ®s
->DmaWriteCfg
);
1343 writel(0, ®s
->MaskInt
);
1344 writel(1, ®s
->IfIdx
);
1347 * McKinley boxes do not like us fiddling with AssistState
1350 writel(1, ®s
->AssistState
);
1353 writel(DEF_STAT
, ®s
->TuneStatTicks
);
1354 writel(DEF_TRACE
, ®s
->TuneTrace
);
1356 ace_set_rxtx_parms(dev
, 0);
1358 if (board_idx
== BOARD_IDX_OVERFLOW
) {
1359 printk(KERN_WARNING
"%s: more than %i NICs detected, "
1360 "ignoring module parameters!\n",
1361 ap
->name
, ACE_MAX_MOD_PARMS
);
1362 } else if (board_idx
>= 0) {
1363 if (tx_coal_tick
[board_idx
])
1364 writel(tx_coal_tick
[board_idx
],
1365 ®s
->TuneTxCoalTicks
);
1366 if (max_tx_desc
[board_idx
])
1367 writel(max_tx_desc
[board_idx
], ®s
->TuneMaxTxDesc
);
1369 if (rx_coal_tick
[board_idx
])
1370 writel(rx_coal_tick
[board_idx
],
1371 ®s
->TuneRxCoalTicks
);
1372 if (max_rx_desc
[board_idx
])
1373 writel(max_rx_desc
[board_idx
], ®s
->TuneMaxRxDesc
);
1375 if (trace
[board_idx
])
1376 writel(trace
[board_idx
], ®s
->TuneTrace
);
1378 if ((tx_ratio
[board_idx
] > 0) && (tx_ratio
[board_idx
] < 64))
1379 writel(tx_ratio
[board_idx
], ®s
->TxBufRat
);
1383 * Default link parameters
1385 tmp
= LNK_ENABLE
| LNK_FULL_DUPLEX
| LNK_1000MB
| LNK_100MB
|
1386 LNK_10MB
| LNK_RX_FLOW_CTL_Y
| LNK_NEG_FCTL
| LNK_NEGOTIATE
;
1387 if(ap
->version
>= 2)
1388 tmp
|= LNK_TX_FLOW_CTL_Y
;
1391 * Override link default parameters
1393 if ((board_idx
>= 0) && link_state
[board_idx
]) {
1394 int option
= link_state
[board_idx
];
1398 if (option
& 0x01) {
1399 printk(KERN_INFO
"%s: Setting half duplex link\n",
1401 tmp
&= ~LNK_FULL_DUPLEX
;
1404 tmp
&= ~LNK_NEGOTIATE
;
1411 if ((option
& 0x70) == 0) {
1412 printk(KERN_WARNING
"%s: No media speed specified, "
1413 "forcing auto negotiation\n", ap
->name
);
1414 tmp
|= LNK_NEGOTIATE
| LNK_1000MB
|
1415 LNK_100MB
| LNK_10MB
;
1417 if ((option
& 0x100) == 0)
1418 tmp
|= LNK_NEG_FCTL
;
1420 printk(KERN_INFO
"%s: Disabling flow control "
1421 "negotiation\n", ap
->name
);
1423 tmp
|= LNK_RX_FLOW_CTL_Y
;
1424 if ((option
& 0x400) && (ap
->version
>= 2)) {
1425 printk(KERN_INFO
"%s: Enabling TX flow control\n",
1427 tmp
|= LNK_TX_FLOW_CTL_Y
;
1432 writel(tmp
, ®s
->TuneLink
);
1433 if (ap
->version
>= 2)
1434 writel(tmp
, ®s
->TuneFastLink
);
1436 writel(ap
->firmware_start
, ®s
->Pc
);
1438 writel(0, ®s
->Mb0Lo
);
1441 * Set tx_csm before we start receiving interrupts, otherwise
1442 * the interrupt handler might think it is supposed to process
1443 * tx ints before we are up and running, which may cause a null
1444 * pointer access in the int handler.
1447 ap
->tx_prd
= *(ap
->tx_csm
) = ap
->tx_ret_csm
= 0;
1450 ace_set_txprd(regs
, ap
, 0);
1451 writel(0, ®s
->RxRetCsm
);
1454 * Enable DMA engine now.
1455 * If we do this sooner, Mckinley box pukes.
1456 * I assume it's because Tigon II DMA engine wants to check
1457 * *something* even before the CPU is started.
1459 writel(1, ®s
->AssistState
); /* enable DMA */
1464 writel(readl(®s
->CpuCtrl
) & ~(CPU_HALT
|CPU_TRACE
), ®s
->CpuCtrl
);
1465 readl(®s
->CpuCtrl
);
1468 * Wait for the firmware to spin up - max 3 seconds.
1470 myjif
= jiffies
+ 3 * HZ
;
1471 while (time_before(jiffies
, myjif
) && !ap
->fw_running
)
1474 if (!ap
->fw_running
) {
1475 printk(KERN_ERR
"%s: Firmware NOT running!\n", ap
->name
);
1478 writel(readl(®s
->CpuCtrl
) | CPU_HALT
, ®s
->CpuCtrl
);
1479 readl(®s
->CpuCtrl
);
1481 /* aman@sgi.com - account for badly behaving firmware/NIC:
1482 * - have observed that the NIC may continue to generate
1483 * interrupts for some reason; attempt to stop it - halt
1484 * second CPU for Tigon II cards, and also clear Mb0
1485 * - if we're a module, we'll fail to load if this was
1486 * the only GbE card in the system => if the kernel does
1487 * see an interrupt from the NIC, code to handle it is
1488 * gone and OOps! - so free_irq also
1490 if (ap
->version
>= 2)
1491 writel(readl(®s
->CpuBCtrl
) | CPU_HALT
,
1493 writel(0, ®s
->Mb0Lo
);
1494 readl(®s
->Mb0Lo
);
1501 * We load the ring here as there seem to be no way to tell the
1502 * firmware to wipe the ring without re-initializing it.
1504 if (!test_and_set_bit(0, &ap
->std_refill_busy
))
1505 ace_load_std_rx_ring(dev
, RX_RING_SIZE
);
1507 printk(KERN_ERR
"%s: Someone is busy refilling the RX ring\n",
1509 if (ap
->version
>= 2) {
1510 if (!test_and_set_bit(0, &ap
->mini_refill_busy
))
1511 ace_load_mini_rx_ring(dev
, RX_MINI_SIZE
);
1513 printk(KERN_ERR
"%s: Someone is busy refilling "
1514 "the RX mini ring\n", ap
->name
);
1519 ace_init_cleanup(dev
);
1524 static void ace_set_rxtx_parms(struct net_device
*dev
, int jumbo
)
1526 struct ace_private
*ap
= netdev_priv(dev
);
1527 struct ace_regs __iomem
*regs
= ap
->regs
;
1528 int board_idx
= ap
->board_idx
;
1530 if (board_idx
>= 0) {
1532 if (!tx_coal_tick
[board_idx
])
1533 writel(DEF_TX_COAL
, ®s
->TuneTxCoalTicks
);
1534 if (!max_tx_desc
[board_idx
])
1535 writel(DEF_TX_MAX_DESC
, ®s
->TuneMaxTxDesc
);
1536 if (!rx_coal_tick
[board_idx
])
1537 writel(DEF_RX_COAL
, ®s
->TuneRxCoalTicks
);
1538 if (!max_rx_desc
[board_idx
])
1539 writel(DEF_RX_MAX_DESC
, ®s
->TuneMaxRxDesc
);
1540 if (!tx_ratio
[board_idx
])
1541 writel(DEF_TX_RATIO
, ®s
->TxBufRat
);
1543 if (!tx_coal_tick
[board_idx
])
1544 writel(DEF_JUMBO_TX_COAL
,
1545 ®s
->TuneTxCoalTicks
);
1546 if (!max_tx_desc
[board_idx
])
1547 writel(DEF_JUMBO_TX_MAX_DESC
,
1548 ®s
->TuneMaxTxDesc
);
1549 if (!rx_coal_tick
[board_idx
])
1550 writel(DEF_JUMBO_RX_COAL
,
1551 ®s
->TuneRxCoalTicks
);
1552 if (!max_rx_desc
[board_idx
])
1553 writel(DEF_JUMBO_RX_MAX_DESC
,
1554 ®s
->TuneMaxRxDesc
);
1555 if (!tx_ratio
[board_idx
])
1556 writel(DEF_JUMBO_TX_RATIO
, ®s
->TxBufRat
);
1562 static void ace_watchdog(struct net_device
*data
)
1564 struct net_device
*dev
= data
;
1565 struct ace_private
*ap
= netdev_priv(dev
);
1566 struct ace_regs __iomem
*regs
= ap
->regs
;
1569 * We haven't received a stats update event for more than 2.5
1570 * seconds and there is data in the transmit queue, thus we
1571 * assume the card is stuck.
1573 if (*ap
->tx_csm
!= ap
->tx_ret_csm
) {
1574 printk(KERN_WARNING
"%s: Transmitter is stuck, %08x\n",
1575 dev
->name
, (unsigned int)readl(®s
->HostCtrl
));
1576 /* This can happen due to ieee flow control. */
1578 printk(KERN_DEBUG
"%s: BUG... transmitter died. Kicking it.\n",
1581 netif_wake_queue(dev
);
1587 static void ace_tasklet(unsigned long arg
)
1589 struct net_device
*dev
= (struct net_device
*) arg
;
1590 struct ace_private
*ap
= netdev_priv(dev
);
1593 cur_size
= atomic_read(&ap
->cur_rx_bufs
);
1594 if ((cur_size
< RX_LOW_STD_THRES
) &&
1595 !test_and_set_bit(0, &ap
->std_refill_busy
)) {
1597 printk("refilling buffers (current %i)\n", cur_size
);
1599 ace_load_std_rx_ring(dev
, RX_RING_SIZE
- cur_size
);
1602 if (ap
->version
>= 2) {
1603 cur_size
= atomic_read(&ap
->cur_mini_bufs
);
1604 if ((cur_size
< RX_LOW_MINI_THRES
) &&
1605 !test_and_set_bit(0, &ap
->mini_refill_busy
)) {
1607 printk("refilling mini buffers (current %i)\n",
1610 ace_load_mini_rx_ring(dev
, RX_MINI_SIZE
- cur_size
);
1614 cur_size
= atomic_read(&ap
->cur_jumbo_bufs
);
1615 if (ap
->jumbo
&& (cur_size
< RX_LOW_JUMBO_THRES
) &&
1616 !test_and_set_bit(0, &ap
->jumbo_refill_busy
)) {
1618 printk("refilling jumbo buffers (current %i)\n", cur_size
);
1620 ace_load_jumbo_rx_ring(dev
, RX_JUMBO_SIZE
- cur_size
);
1622 ap
->tasklet_pending
= 0;
1627 * Copy the contents of the NIC's trace buffer to kernel memory.
1629 static void ace_dump_trace(struct ace_private
*ap
)
1633 if (!(ap
->trace_buf
= kmalloc(ACE_TRACE_SIZE
, GFP_KERNEL
)))
1640 * Load the standard rx ring.
1642 * Loading rings is safe without holding the spin lock since this is
1643 * done only before the device is enabled, thus no interrupts are
1644 * generated and by the interrupt handler/tasklet handler.
1646 static void ace_load_std_rx_ring(struct net_device
*dev
, int nr_bufs
)
1648 struct ace_private
*ap
= netdev_priv(dev
);
1649 struct ace_regs __iomem
*regs
= ap
->regs
;
1653 prefetchw(&ap
->cur_rx_bufs
);
1655 idx
= ap
->rx_std_skbprd
;
1657 for (i
= 0; i
< nr_bufs
; i
++) {
1658 struct sk_buff
*skb
;
1662 skb
= netdev_alloc_skb_ip_align(dev
, ACE_STD_BUFSIZE
);
1666 mapping
= pci_map_page(ap
->pdev
, virt_to_page(skb
->data
),
1667 offset_in_page(skb
->data
),
1669 PCI_DMA_FROMDEVICE
);
1670 ap
->skb
->rx_std_skbuff
[idx
].skb
= skb
;
1671 dma_unmap_addr_set(&ap
->skb
->rx_std_skbuff
[idx
],
1674 rd
= &ap
->rx_std_ring
[idx
];
1675 set_aceaddr(&rd
->addr
, mapping
);
1676 rd
->size
= ACE_STD_BUFSIZE
;
1678 idx
= (idx
+ 1) % RX_STD_RING_ENTRIES
;
1684 atomic_add(i
, &ap
->cur_rx_bufs
);
1685 ap
->rx_std_skbprd
= idx
;
1687 if (ACE_IS_TIGON_I(ap
)) {
1689 cmd
.evt
= C_SET_RX_PRD_IDX
;
1691 cmd
.idx
= ap
->rx_std_skbprd
;
1692 ace_issue_cmd(regs
, &cmd
);
1694 writel(idx
, ®s
->RxStdPrd
);
1699 clear_bit(0, &ap
->std_refill_busy
);
1703 printk(KERN_INFO
"Out of memory when allocating "
1704 "standard receive buffers\n");
1709 static void ace_load_mini_rx_ring(struct net_device
*dev
, int nr_bufs
)
1711 struct ace_private
*ap
= netdev_priv(dev
);
1712 struct ace_regs __iomem
*regs
= ap
->regs
;
1715 prefetchw(&ap
->cur_mini_bufs
);
1717 idx
= ap
->rx_mini_skbprd
;
1718 for (i
= 0; i
< nr_bufs
; i
++) {
1719 struct sk_buff
*skb
;
1723 skb
= netdev_alloc_skb_ip_align(dev
, ACE_MINI_BUFSIZE
);
1727 mapping
= pci_map_page(ap
->pdev
, virt_to_page(skb
->data
),
1728 offset_in_page(skb
->data
),
1730 PCI_DMA_FROMDEVICE
);
1731 ap
->skb
->rx_mini_skbuff
[idx
].skb
= skb
;
1732 dma_unmap_addr_set(&ap
->skb
->rx_mini_skbuff
[idx
],
1735 rd
= &ap
->rx_mini_ring
[idx
];
1736 set_aceaddr(&rd
->addr
, mapping
);
1737 rd
->size
= ACE_MINI_BUFSIZE
;
1739 idx
= (idx
+ 1) % RX_MINI_RING_ENTRIES
;
1745 atomic_add(i
, &ap
->cur_mini_bufs
);
1747 ap
->rx_mini_skbprd
= idx
;
1749 writel(idx
, ®s
->RxMiniPrd
);
1753 clear_bit(0, &ap
->mini_refill_busy
);
1756 printk(KERN_INFO
"Out of memory when allocating "
1757 "mini receive buffers\n");
1763 * Load the jumbo rx ring, this may happen at any time if the MTU
1764 * is changed to a value > 1500.
1766 static void ace_load_jumbo_rx_ring(struct net_device
*dev
, int nr_bufs
)
1768 struct ace_private
*ap
= netdev_priv(dev
);
1769 struct ace_regs __iomem
*regs
= ap
->regs
;
1772 idx
= ap
->rx_jumbo_skbprd
;
1774 for (i
= 0; i
< nr_bufs
; i
++) {
1775 struct sk_buff
*skb
;
1779 skb
= netdev_alloc_skb_ip_align(dev
, ACE_JUMBO_BUFSIZE
);
1783 mapping
= pci_map_page(ap
->pdev
, virt_to_page(skb
->data
),
1784 offset_in_page(skb
->data
),
1786 PCI_DMA_FROMDEVICE
);
1787 ap
->skb
->rx_jumbo_skbuff
[idx
].skb
= skb
;
1788 dma_unmap_addr_set(&ap
->skb
->rx_jumbo_skbuff
[idx
],
1791 rd
= &ap
->rx_jumbo_ring
[idx
];
1792 set_aceaddr(&rd
->addr
, mapping
);
1793 rd
->size
= ACE_JUMBO_BUFSIZE
;
1795 idx
= (idx
+ 1) % RX_JUMBO_RING_ENTRIES
;
1801 atomic_add(i
, &ap
->cur_jumbo_bufs
);
1802 ap
->rx_jumbo_skbprd
= idx
;
1804 if (ACE_IS_TIGON_I(ap
)) {
1806 cmd
.evt
= C_SET_RX_JUMBO_PRD_IDX
;
1808 cmd
.idx
= ap
->rx_jumbo_skbprd
;
1809 ace_issue_cmd(regs
, &cmd
);
1811 writel(idx
, ®s
->RxJumboPrd
);
1816 clear_bit(0, &ap
->jumbo_refill_busy
);
1819 if (net_ratelimit())
1820 printk(KERN_INFO
"Out of memory when allocating "
1821 "jumbo receive buffers\n");
1827 * All events are considered to be slow (RX/TX ints do not generate
1828 * events) and are handled here, outside the main interrupt handler,
1829 * to reduce the size of the handler.
1831 static u32
ace_handle_event(struct net_device
*dev
, u32 evtcsm
, u32 evtprd
)
1833 struct ace_private
*ap
;
1835 ap
= netdev_priv(dev
);
1837 while (evtcsm
!= evtprd
) {
1838 switch (ap
->evt_ring
[evtcsm
].evt
) {
1840 printk(KERN_INFO
"%s: Firmware up and running\n",
1845 case E_STATS_UPDATED
:
1849 u16 code
= ap
->evt_ring
[evtcsm
].code
;
1853 u32 state
= readl(&ap
->regs
->GigLnkState
);
1854 printk(KERN_WARNING
"%s: Optical link UP "
1855 "(%s Duplex, Flow Control: %s%s)\n",
1857 state
& LNK_FULL_DUPLEX
? "Full":"Half",
1858 state
& LNK_TX_FLOW_CTL_Y
? "TX " : "",
1859 state
& LNK_RX_FLOW_CTL_Y
? "RX" : "");
1863 printk(KERN_WARNING
"%s: Optical link DOWN\n",
1866 case E_C_LINK_10_100
:
1867 printk(KERN_WARNING
"%s: 10/100BaseT link "
1871 printk(KERN_ERR
"%s: Unknown optical link "
1872 "state %02x\n", ap
->name
, code
);
1877 switch(ap
->evt_ring
[evtcsm
].code
) {
1878 case E_C_ERR_INVAL_CMD
:
1879 printk(KERN_ERR
"%s: invalid command error\n",
1882 case E_C_ERR_UNIMP_CMD
:
1883 printk(KERN_ERR
"%s: unimplemented command "
1884 "error\n", ap
->name
);
1886 case E_C_ERR_BAD_CFG
:
1887 printk(KERN_ERR
"%s: bad config error\n",
1891 printk(KERN_ERR
"%s: unknown error %02x\n",
1892 ap
->name
, ap
->evt_ring
[evtcsm
].code
);
1895 case E_RESET_JUMBO_RNG
:
1898 for (i
= 0; i
< RX_JUMBO_RING_ENTRIES
; i
++) {
1899 if (ap
->skb
->rx_jumbo_skbuff
[i
].skb
) {
1900 ap
->rx_jumbo_ring
[i
].size
= 0;
1901 set_aceaddr(&ap
->rx_jumbo_ring
[i
].addr
, 0);
1902 dev_kfree_skb(ap
->skb
->rx_jumbo_skbuff
[i
].skb
);
1903 ap
->skb
->rx_jumbo_skbuff
[i
].skb
= NULL
;
1907 if (ACE_IS_TIGON_I(ap
)) {
1909 cmd
.evt
= C_SET_RX_JUMBO_PRD_IDX
;
1912 ace_issue_cmd(ap
->regs
, &cmd
);
1914 writel(0, &((ap
->regs
)->RxJumboPrd
));
1919 ap
->rx_jumbo_skbprd
= 0;
1920 printk(KERN_INFO
"%s: Jumbo ring flushed\n",
1922 clear_bit(0, &ap
->jumbo_refill_busy
);
1926 printk(KERN_ERR
"%s: Unhandled event 0x%02x\n",
1927 ap
->name
, ap
->evt_ring
[evtcsm
].evt
);
1929 evtcsm
= (evtcsm
+ 1) % EVT_RING_ENTRIES
;
1936 static void ace_rx_int(struct net_device
*dev
, u32 rxretprd
, u32 rxretcsm
)
1938 struct ace_private
*ap
= netdev_priv(dev
);
1940 int mini_count
= 0, std_count
= 0;
1944 prefetchw(&ap
->cur_rx_bufs
);
1945 prefetchw(&ap
->cur_mini_bufs
);
1947 while (idx
!= rxretprd
) {
1948 struct ring_info
*rip
;
1949 struct sk_buff
*skb
;
1950 struct rx_desc
*rxdesc
, *retdesc
;
1952 int bd_flags
, desc_type
, mapsize
;
1956 /* make sure the rx descriptor isn't read before rxretprd */
1957 if (idx
== rxretcsm
)
1960 retdesc
= &ap
->rx_return_ring
[idx
];
1961 skbidx
= retdesc
->idx
;
1962 bd_flags
= retdesc
->flags
;
1963 desc_type
= bd_flags
& (BD_FLG_JUMBO
| BD_FLG_MINI
);
1967 * Normal frames do not have any flags set
1969 * Mini and normal frames arrive frequently,
1970 * so use a local counter to avoid doing
1971 * atomic operations for each packet arriving.
1974 rip
= &ap
->skb
->rx_std_skbuff
[skbidx
];
1975 mapsize
= ACE_STD_BUFSIZE
;
1976 rxdesc
= &ap
->rx_std_ring
[skbidx
];
1980 rip
= &ap
->skb
->rx_jumbo_skbuff
[skbidx
];
1981 mapsize
= ACE_JUMBO_BUFSIZE
;
1982 rxdesc
= &ap
->rx_jumbo_ring
[skbidx
];
1983 atomic_dec(&ap
->cur_jumbo_bufs
);
1986 rip
= &ap
->skb
->rx_mini_skbuff
[skbidx
];
1987 mapsize
= ACE_MINI_BUFSIZE
;
1988 rxdesc
= &ap
->rx_mini_ring
[skbidx
];
1992 printk(KERN_INFO
"%s: unknown frame type (0x%02x) "
1993 "returned by NIC\n", dev
->name
,
2000 pci_unmap_page(ap
->pdev
,
2001 dma_unmap_addr(rip
, mapping
),
2003 PCI_DMA_FROMDEVICE
);
2004 skb_put(skb
, retdesc
->size
);
2009 csum
= retdesc
->tcp_udp_csum
;
2011 skb
->protocol
= eth_type_trans(skb
, dev
);
2014 * Instead of forcing the poor tigon mips cpu to calculate
2015 * pseudo hdr checksum, we do this ourselves.
2017 if (bd_flags
& BD_FLG_TCP_UDP_SUM
) {
2018 skb
->csum
= htons(csum
);
2019 skb
->ip_summed
= CHECKSUM_COMPLETE
;
2021 skb_checksum_none_assert(skb
);
2025 if ((bd_flags
& BD_FLG_VLAN_TAG
))
2026 __vlan_hwaccel_put_tag(skb
, retdesc
->vlan
);
2029 dev
->stats
.rx_packets
++;
2030 dev
->stats
.rx_bytes
+= retdesc
->size
;
2032 idx
= (idx
+ 1) % RX_RETURN_RING_ENTRIES
;
2035 atomic_sub(std_count
, &ap
->cur_rx_bufs
);
2036 if (!ACE_IS_TIGON_I(ap
))
2037 atomic_sub(mini_count
, &ap
->cur_mini_bufs
);
2041 * According to the documentation RxRetCsm is obsolete with
2042 * the 12.3.x Firmware - my Tigon I NICs seem to disagree!
2044 if (ACE_IS_TIGON_I(ap
)) {
2045 writel(idx
, &ap
->regs
->RxRetCsm
);
2056 static inline void ace_tx_int(struct net_device
*dev
,
2059 struct ace_private
*ap
= netdev_priv(dev
);
2062 struct sk_buff
*skb
;
2063 struct tx_ring_info
*info
;
2065 info
= ap
->skb
->tx_skbuff
+ idx
;
2068 if (dma_unmap_len(info
, maplen
)) {
2069 pci_unmap_page(ap
->pdev
, dma_unmap_addr(info
, mapping
),
2070 dma_unmap_len(info
, maplen
),
2072 dma_unmap_len_set(info
, maplen
, 0);
2076 dev
->stats
.tx_packets
++;
2077 dev
->stats
.tx_bytes
+= skb
->len
;
2078 dev_kfree_skb_irq(skb
);
2082 idx
= (idx
+ 1) % ACE_TX_RING_ENTRIES(ap
);
2083 } while (idx
!= txcsm
);
2085 if (netif_queue_stopped(dev
))
2086 netif_wake_queue(dev
);
2089 ap
->tx_ret_csm
= txcsm
;
2091 /* So... tx_ret_csm is advanced _after_ check for device wakeup.
2093 * We could try to make it before. In this case we would get
2094 * the following race condition: hard_start_xmit on other cpu
2095 * enters after we advanced tx_ret_csm and fills space,
2096 * which we have just freed, so that we make illegal device wakeup.
2097 * There is no good way to workaround this (at entry
2098 * to ace_start_xmit detects this condition and prevents
2099 * ring corruption, but it is not a good workaround.)
2101 * When tx_ret_csm is advanced after, we wake up device _only_
2102 * if we really have some space in ring (though the core doing
2103 * hard_start_xmit can see full ring for some period and has to
2104 * synchronize.) Superb.
2105 * BUT! We get another subtle race condition. hard_start_xmit
2106 * may think that ring is full between wakeup and advancing
2107 * tx_ret_csm and will stop device instantly! It is not so bad.
2108 * We are guaranteed that there is something in ring, so that
2109 * the next irq will resume transmission. To speedup this we could
2110 * mark descriptor, which closes ring with BD_FLG_COAL_NOW
2111 * (see ace_start_xmit).
2113 * Well, this dilemma exists in all lock-free devices.
2114 * We, following scheme used in drivers by Donald Becker,
2115 * select the least dangerous.
2121 static irqreturn_t
ace_interrupt(int irq
, void *dev_id
)
2123 struct net_device
*dev
= (struct net_device
*)dev_id
;
2124 struct ace_private
*ap
= netdev_priv(dev
);
2125 struct ace_regs __iomem
*regs
= ap
->regs
;
2127 u32 txcsm
, rxretcsm
, rxretprd
;
2131 * In case of PCI shared interrupts or spurious interrupts,
2132 * we want to make sure it is actually our interrupt before
2133 * spending any time in here.
2135 if (!(readl(®s
->HostCtrl
) & IN_INT
))
2139 * ACK intr now. Otherwise we will lose updates to rx_ret_prd,
2140 * which happened _after_ rxretprd = *ap->rx_ret_prd; but before
2141 * writel(0, ®s->Mb0Lo).
2143 * "IRQ avoidance" recommended in docs applies to IRQs served
2144 * threads and it is wrong even for that case.
2146 writel(0, ®s
->Mb0Lo
);
2147 readl(®s
->Mb0Lo
);
2150 * There is no conflict between transmit handling in
2151 * start_xmit and receive processing, thus there is no reason
2152 * to take a spin lock for RX handling. Wait until we start
2153 * working on the other stuff - hey we don't need a spin lock
2156 rxretprd
= *ap
->rx_ret_prd
;
2157 rxretcsm
= ap
->cur_rx
;
2159 if (rxretprd
!= rxretcsm
)
2160 ace_rx_int(dev
, rxretprd
, rxretcsm
);
2162 txcsm
= *ap
->tx_csm
;
2163 idx
= ap
->tx_ret_csm
;
2167 * If each skb takes only one descriptor this check degenerates
2168 * to identity, because new space has just been opened.
2169 * But if skbs are fragmented we must check that this index
2170 * update releases enough of space, otherwise we just
2171 * wait for device to make more work.
2173 if (!tx_ring_full(ap
, txcsm
, ap
->tx_prd
))
2174 ace_tx_int(dev
, txcsm
, idx
);
2177 evtcsm
= readl(®s
->EvtCsm
);
2178 evtprd
= *ap
->evt_prd
;
2180 if (evtcsm
!= evtprd
) {
2181 evtcsm
= ace_handle_event(dev
, evtcsm
, evtprd
);
2182 writel(evtcsm
, ®s
->EvtCsm
);
2186 * This has to go last in the interrupt handler and run with
2187 * the spin lock released ... what lock?
2189 if (netif_running(dev
)) {
2191 int run_tasklet
= 0;
2193 cur_size
= atomic_read(&ap
->cur_rx_bufs
);
2194 if (cur_size
< RX_LOW_STD_THRES
) {
2195 if ((cur_size
< RX_PANIC_STD_THRES
) &&
2196 !test_and_set_bit(0, &ap
->std_refill_busy
)) {
2198 printk("low on std buffers %i\n", cur_size
);
2200 ace_load_std_rx_ring(dev
,
2201 RX_RING_SIZE
- cur_size
);
2206 if (!ACE_IS_TIGON_I(ap
)) {
2207 cur_size
= atomic_read(&ap
->cur_mini_bufs
);
2208 if (cur_size
< RX_LOW_MINI_THRES
) {
2209 if ((cur_size
< RX_PANIC_MINI_THRES
) &&
2210 !test_and_set_bit(0,
2211 &ap
->mini_refill_busy
)) {
2213 printk("low on mini buffers %i\n",
2216 ace_load_mini_rx_ring(dev
,
2217 RX_MINI_SIZE
- cur_size
);
2224 cur_size
= atomic_read(&ap
->cur_jumbo_bufs
);
2225 if (cur_size
< RX_LOW_JUMBO_THRES
) {
2226 if ((cur_size
< RX_PANIC_JUMBO_THRES
) &&
2227 !test_and_set_bit(0,
2228 &ap
->jumbo_refill_busy
)){
2230 printk("low on jumbo buffers %i\n",
2233 ace_load_jumbo_rx_ring(dev
,
2234 RX_JUMBO_SIZE
- cur_size
);
2239 if (run_tasklet
&& !ap
->tasklet_pending
) {
2240 ap
->tasklet_pending
= 1;
2241 tasklet_schedule(&ap
->ace_tasklet
);
2248 static int ace_open(struct net_device
*dev
)
2250 struct ace_private
*ap
= netdev_priv(dev
);
2251 struct ace_regs __iomem
*regs
= ap
->regs
;
2254 if (!(ap
->fw_running
)) {
2255 printk(KERN_WARNING
"%s: Firmware not running!\n", dev
->name
);
2259 writel(dev
->mtu
+ ETH_HLEN
+ 4, ®s
->IfMtu
);
2261 cmd
.evt
= C_CLEAR_STATS
;
2264 ace_issue_cmd(regs
, &cmd
);
2266 cmd
.evt
= C_HOST_STATE
;
2267 cmd
.code
= C_C_STACK_UP
;
2269 ace_issue_cmd(regs
, &cmd
);
2272 !test_and_set_bit(0, &ap
->jumbo_refill_busy
))
2273 ace_load_jumbo_rx_ring(dev
, RX_JUMBO_SIZE
);
2275 if (dev
->flags
& IFF_PROMISC
) {
2276 cmd
.evt
= C_SET_PROMISC_MODE
;
2277 cmd
.code
= C_C_PROMISC_ENABLE
;
2279 ace_issue_cmd(regs
, &cmd
);
2287 cmd
.evt
= C_LNK_NEGOTIATION
;
2290 ace_issue_cmd(regs
, &cmd
);
2293 netif_start_queue(dev
);
2296 * Setup the bottom half rx ring refill handler
2298 tasklet_init(&ap
->ace_tasklet
, ace_tasklet
, (unsigned long)dev
);
2303 static int ace_close(struct net_device
*dev
)
2305 struct ace_private
*ap
= netdev_priv(dev
);
2306 struct ace_regs __iomem
*regs
= ap
->regs
;
2308 unsigned long flags
;
2312 * Without (or before) releasing irq and stopping hardware, this
2313 * is an absolute non-sense, by the way. It will be reset instantly
2316 netif_stop_queue(dev
);
2320 cmd
.evt
= C_SET_PROMISC_MODE
;
2321 cmd
.code
= C_C_PROMISC_DISABLE
;
2323 ace_issue_cmd(regs
, &cmd
);
2327 cmd
.evt
= C_HOST_STATE
;
2328 cmd
.code
= C_C_STACK_DOWN
;
2330 ace_issue_cmd(regs
, &cmd
);
2332 tasklet_kill(&ap
->ace_tasklet
);
2335 * Make sure one CPU is not processing packets while
2336 * buffers are being released by another.
2339 local_irq_save(flags
);
2342 for (i
= 0; i
< ACE_TX_RING_ENTRIES(ap
); i
++) {
2343 struct sk_buff
*skb
;
2344 struct tx_ring_info
*info
;
2346 info
= ap
->skb
->tx_skbuff
+ i
;
2349 if (dma_unmap_len(info
, maplen
)) {
2350 if (ACE_IS_TIGON_I(ap
)) {
2351 /* NB: TIGON_1 is special, tx_ring is in io space */
2352 struct tx_desc __iomem
*tx
;
2353 tx
= (__force
struct tx_desc __iomem
*) &ap
->tx_ring
[i
];
2354 writel(0, &tx
->addr
.addrhi
);
2355 writel(0, &tx
->addr
.addrlo
);
2356 writel(0, &tx
->flagsize
);
2358 memset(ap
->tx_ring
+ i
, 0,
2359 sizeof(struct tx_desc
));
2360 pci_unmap_page(ap
->pdev
, dma_unmap_addr(info
, mapping
),
2361 dma_unmap_len(info
, maplen
),
2363 dma_unmap_len_set(info
, maplen
, 0);
2372 cmd
.evt
= C_RESET_JUMBO_RNG
;
2375 ace_issue_cmd(regs
, &cmd
);
2378 ace_unmask_irq(dev
);
2379 local_irq_restore(flags
);
2385 static inline dma_addr_t
2386 ace_map_tx_skb(struct ace_private
*ap
, struct sk_buff
*skb
,
2387 struct sk_buff
*tail
, u32 idx
)
2390 struct tx_ring_info
*info
;
2392 mapping
= pci_map_page(ap
->pdev
, virt_to_page(skb
->data
),
2393 offset_in_page(skb
->data
),
2394 skb
->len
, PCI_DMA_TODEVICE
);
2396 info
= ap
->skb
->tx_skbuff
+ idx
;
2398 dma_unmap_addr_set(info
, mapping
, mapping
);
2399 dma_unmap_len_set(info
, maplen
, skb
->len
);
2405 ace_load_tx_bd(struct ace_private
*ap
, struct tx_desc
*desc
, u64 addr
,
2406 u32 flagsize
, u32 vlan_tag
)
2408 #if !USE_TX_COAL_NOW
2409 flagsize
&= ~BD_FLG_COAL_NOW
;
2412 if (ACE_IS_TIGON_I(ap
)) {
2413 struct tx_desc __iomem
*io
= (__force
struct tx_desc __iomem
*) desc
;
2414 writel(addr
>> 32, &io
->addr
.addrhi
);
2415 writel(addr
& 0xffffffff, &io
->addr
.addrlo
);
2416 writel(flagsize
, &io
->flagsize
);
2417 writel(vlan_tag
, &io
->vlanres
);
2419 desc
->addr
.addrhi
= addr
>> 32;
2420 desc
->addr
.addrlo
= addr
;
2421 desc
->flagsize
= flagsize
;
2422 desc
->vlanres
= vlan_tag
;
2427 static netdev_tx_t
ace_start_xmit(struct sk_buff
*skb
,
2428 struct net_device
*dev
)
2430 struct ace_private
*ap
= netdev_priv(dev
);
2431 struct ace_regs __iomem
*regs
= ap
->regs
;
2432 struct tx_desc
*desc
;
2434 unsigned long maxjiff
= jiffies
+ 3*HZ
;
2439 if (tx_ring_full(ap
, ap
->tx_ret_csm
, idx
))
2442 if (!skb_shinfo(skb
)->nr_frags
) {
2446 mapping
= ace_map_tx_skb(ap
, skb
, skb
, idx
);
2447 flagsize
= (skb
->len
<< 16) | (BD_FLG_END
);
2448 if (skb
->ip_summed
== CHECKSUM_PARTIAL
)
2449 flagsize
|= BD_FLG_TCP_UDP_SUM
;
2450 if (vlan_tx_tag_present(skb
)) {
2451 flagsize
|= BD_FLG_VLAN_TAG
;
2452 vlan_tag
= vlan_tx_tag_get(skb
);
2454 desc
= ap
->tx_ring
+ idx
;
2455 idx
= (idx
+ 1) % ACE_TX_RING_ENTRIES(ap
);
2457 /* Look at ace_tx_int for explanations. */
2458 if (tx_ring_full(ap
, ap
->tx_ret_csm
, idx
))
2459 flagsize
|= BD_FLG_COAL_NOW
;
2461 ace_load_tx_bd(ap
, desc
, mapping
, flagsize
, vlan_tag
);
2467 mapping
= ace_map_tx_skb(ap
, skb
, NULL
, idx
);
2468 flagsize
= (skb_headlen(skb
) << 16);
2469 if (skb
->ip_summed
== CHECKSUM_PARTIAL
)
2470 flagsize
|= BD_FLG_TCP_UDP_SUM
;
2471 if (vlan_tx_tag_present(skb
)) {
2472 flagsize
|= BD_FLG_VLAN_TAG
;
2473 vlan_tag
= vlan_tx_tag_get(skb
);
2476 ace_load_tx_bd(ap
, ap
->tx_ring
+ idx
, mapping
, flagsize
, vlan_tag
);
2478 idx
= (idx
+ 1) % ACE_TX_RING_ENTRIES(ap
);
2480 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
2481 skb_frag_t
*frag
= &skb_shinfo(skb
)->frags
[i
];
2482 struct tx_ring_info
*info
;
2485 info
= ap
->skb
->tx_skbuff
+ idx
;
2486 desc
= ap
->tx_ring
+ idx
;
2488 mapping
= skb_frag_dma_map(&ap
->pdev
->dev
, frag
, 0,
2492 flagsize
= (frag
->size
<< 16);
2493 if (skb
->ip_summed
== CHECKSUM_PARTIAL
)
2494 flagsize
|= BD_FLG_TCP_UDP_SUM
;
2495 idx
= (idx
+ 1) % ACE_TX_RING_ENTRIES(ap
);
2497 if (i
== skb_shinfo(skb
)->nr_frags
- 1) {
2498 flagsize
|= BD_FLG_END
;
2499 if (tx_ring_full(ap
, ap
->tx_ret_csm
, idx
))
2500 flagsize
|= BD_FLG_COAL_NOW
;
2503 * Only the last fragment frees
2510 dma_unmap_addr_set(info
, mapping
, mapping
);
2511 dma_unmap_len_set(info
, maplen
, frag
->size
);
2512 ace_load_tx_bd(ap
, desc
, mapping
, flagsize
, vlan_tag
);
2518 ace_set_txprd(regs
, ap
, idx
);
2520 if (flagsize
& BD_FLG_COAL_NOW
) {
2521 netif_stop_queue(dev
);
2524 * A TX-descriptor producer (an IRQ) might have gotten
2525 * between, making the ring free again. Since xmit is
2526 * serialized, this is the only situation we have to
2529 if (!tx_ring_full(ap
, ap
->tx_ret_csm
, idx
))
2530 netif_wake_queue(dev
);
2533 return NETDEV_TX_OK
;
2537 * This race condition is unavoidable with lock-free drivers.
2538 * We wake up the queue _before_ tx_prd is advanced, so that we can
2539 * enter hard_start_xmit too early, while tx ring still looks closed.
2540 * This happens ~1-4 times per 100000 packets, so that we can allow
2541 * to loop syncing to other CPU. Probably, we need an additional
2542 * wmb() in ace_tx_intr as well.
2544 * Note that this race is relieved by reserving one more entry
2545 * in tx ring than it is necessary (see original non-SG driver).
2546 * However, with SG we need to reserve 2*MAX_SKB_FRAGS+1, which
2547 * is already overkill.
2549 * Alternative is to return with 1 not throttling queue. In this
2550 * case loop becomes longer, no more useful effects.
2552 if (time_before(jiffies
, maxjiff
)) {
2558 /* The ring is stuck full. */
2559 printk(KERN_WARNING
"%s: Transmit ring stuck full\n", dev
->name
);
2560 return NETDEV_TX_BUSY
;
2564 static int ace_change_mtu(struct net_device
*dev
, int new_mtu
)
2566 struct ace_private
*ap
= netdev_priv(dev
);
2567 struct ace_regs __iomem
*regs
= ap
->regs
;
2569 if (new_mtu
> ACE_JUMBO_MTU
)
2572 writel(new_mtu
+ ETH_HLEN
+ 4, ®s
->IfMtu
);
2575 if (new_mtu
> ACE_STD_MTU
) {
2577 printk(KERN_INFO
"%s: Enabling Jumbo frame "
2578 "support\n", dev
->name
);
2580 if (!test_and_set_bit(0, &ap
->jumbo_refill_busy
))
2581 ace_load_jumbo_rx_ring(dev
, RX_JUMBO_SIZE
);
2582 ace_set_rxtx_parms(dev
, 1);
2585 while (test_and_set_bit(0, &ap
->jumbo_refill_busy
));
2586 ace_sync_irq(dev
->irq
);
2587 ace_set_rxtx_parms(dev
, 0);
2591 cmd
.evt
= C_RESET_JUMBO_RNG
;
2594 ace_issue_cmd(regs
, &cmd
);
2601 static int ace_get_settings(struct net_device
*dev
, struct ethtool_cmd
*ecmd
)
2603 struct ace_private
*ap
= netdev_priv(dev
);
2604 struct ace_regs __iomem
*regs
= ap
->regs
;
2607 memset(ecmd
, 0, sizeof(struct ethtool_cmd
));
2609 (SUPPORTED_10baseT_Half
| SUPPORTED_10baseT_Full
|
2610 SUPPORTED_100baseT_Half
| SUPPORTED_100baseT_Full
|
2611 SUPPORTED_1000baseT_Half
| SUPPORTED_1000baseT_Full
|
2612 SUPPORTED_Autoneg
| SUPPORTED_FIBRE
);
2614 ecmd
->port
= PORT_FIBRE
;
2615 ecmd
->transceiver
= XCVR_INTERNAL
;
2617 link
= readl(®s
->GigLnkState
);
2618 if (link
& LNK_1000MB
)
2619 ethtool_cmd_speed_set(ecmd
, SPEED_1000
);
2621 link
= readl(®s
->FastLnkState
);
2622 if (link
& LNK_100MB
)
2623 ethtool_cmd_speed_set(ecmd
, SPEED_100
);
2624 else if (link
& LNK_10MB
)
2625 ethtool_cmd_speed_set(ecmd
, SPEED_10
);
2627 ethtool_cmd_speed_set(ecmd
, 0);
2629 if (link
& LNK_FULL_DUPLEX
)
2630 ecmd
->duplex
= DUPLEX_FULL
;
2632 ecmd
->duplex
= DUPLEX_HALF
;
2634 if (link
& LNK_NEGOTIATE
)
2635 ecmd
->autoneg
= AUTONEG_ENABLE
;
2637 ecmd
->autoneg
= AUTONEG_DISABLE
;
2641 * Current struct ethtool_cmd is insufficient
2643 ecmd
->trace
= readl(®s
->TuneTrace
);
2645 ecmd
->txcoal
= readl(®s
->TuneTxCoalTicks
);
2646 ecmd
->rxcoal
= readl(®s
->TuneRxCoalTicks
);
2648 ecmd
->maxtxpkt
= readl(®s
->TuneMaxTxDesc
);
2649 ecmd
->maxrxpkt
= readl(®s
->TuneMaxRxDesc
);
2654 static int ace_set_settings(struct net_device
*dev
, struct ethtool_cmd
*ecmd
)
2656 struct ace_private
*ap
= netdev_priv(dev
);
2657 struct ace_regs __iomem
*regs
= ap
->regs
;
2660 link
= readl(®s
->GigLnkState
);
2661 if (link
& LNK_1000MB
)
2664 link
= readl(®s
->FastLnkState
);
2665 if (link
& LNK_100MB
)
2667 else if (link
& LNK_10MB
)
2673 link
= LNK_ENABLE
| LNK_1000MB
| LNK_100MB
| LNK_10MB
|
2674 LNK_RX_FLOW_CTL_Y
| LNK_NEG_FCTL
;
2675 if (!ACE_IS_TIGON_I(ap
))
2676 link
|= LNK_TX_FLOW_CTL_Y
;
2677 if (ecmd
->autoneg
== AUTONEG_ENABLE
)
2678 link
|= LNK_NEGOTIATE
;
2679 if (ethtool_cmd_speed(ecmd
) != speed
) {
2680 link
&= ~(LNK_1000MB
| LNK_100MB
| LNK_10MB
);
2681 switch (ethtool_cmd_speed(ecmd
)) {
2694 if (ecmd
->duplex
== DUPLEX_FULL
)
2695 link
|= LNK_FULL_DUPLEX
;
2697 if (link
!= ap
->link
) {
2699 printk(KERN_INFO
"%s: Renegotiating link state\n",
2703 writel(link
, ®s
->TuneLink
);
2704 if (!ACE_IS_TIGON_I(ap
))
2705 writel(link
, ®s
->TuneFastLink
);
2708 cmd
.evt
= C_LNK_NEGOTIATION
;
2711 ace_issue_cmd(regs
, &cmd
);
2716 static void ace_get_drvinfo(struct net_device
*dev
,
2717 struct ethtool_drvinfo
*info
)
2719 struct ace_private
*ap
= netdev_priv(dev
);
2721 strlcpy(info
->driver
, "acenic", sizeof(info
->driver
));
2722 snprintf(info
->version
, sizeof(info
->version
), "%i.%i.%i",
2723 ap
->firmware_major
, ap
->firmware_minor
,
2727 strlcpy(info
->bus_info
, pci_name(ap
->pdev
),
2728 sizeof(info
->bus_info
));
2733 * Set the hardware MAC address.
2735 static int ace_set_mac_addr(struct net_device
*dev
, void *p
)
2737 struct ace_private
*ap
= netdev_priv(dev
);
2738 struct ace_regs __iomem
*regs
= ap
->regs
;
2739 struct sockaddr
*addr
=p
;
2743 if(netif_running(dev
))
2746 memcpy(dev
->dev_addr
, addr
->sa_data
,dev
->addr_len
);
2748 da
= (u8
*)dev
->dev_addr
;
2750 writel(da
[0] << 8 | da
[1], ®s
->MacAddrHi
);
2751 writel((da
[2] << 24) | (da
[3] << 16) | (da
[4] << 8) | da
[5],
2754 cmd
.evt
= C_SET_MAC_ADDR
;
2757 ace_issue_cmd(regs
, &cmd
);
2763 static void ace_set_multicast_list(struct net_device
*dev
)
2765 struct ace_private
*ap
= netdev_priv(dev
);
2766 struct ace_regs __iomem
*regs
= ap
->regs
;
2769 if ((dev
->flags
& IFF_ALLMULTI
) && !(ap
->mcast_all
)) {
2770 cmd
.evt
= C_SET_MULTICAST_MODE
;
2771 cmd
.code
= C_C_MCAST_ENABLE
;
2773 ace_issue_cmd(regs
, &cmd
);
2775 } else if (ap
->mcast_all
) {
2776 cmd
.evt
= C_SET_MULTICAST_MODE
;
2777 cmd
.code
= C_C_MCAST_DISABLE
;
2779 ace_issue_cmd(regs
, &cmd
);
2783 if ((dev
->flags
& IFF_PROMISC
) && !(ap
->promisc
)) {
2784 cmd
.evt
= C_SET_PROMISC_MODE
;
2785 cmd
.code
= C_C_PROMISC_ENABLE
;
2787 ace_issue_cmd(regs
, &cmd
);
2789 }else if (!(dev
->flags
& IFF_PROMISC
) && (ap
->promisc
)) {
2790 cmd
.evt
= C_SET_PROMISC_MODE
;
2791 cmd
.code
= C_C_PROMISC_DISABLE
;
2793 ace_issue_cmd(regs
, &cmd
);
2798 * For the time being multicast relies on the upper layers
2799 * filtering it properly. The Firmware does not allow one to
2800 * set the entire multicast list at a time and keeping track of
2801 * it here is going to be messy.
2803 if (!netdev_mc_empty(dev
) && !ap
->mcast_all
) {
2804 cmd
.evt
= C_SET_MULTICAST_MODE
;
2805 cmd
.code
= C_C_MCAST_ENABLE
;
2807 ace_issue_cmd(regs
, &cmd
);
2808 }else if (!ap
->mcast_all
) {
2809 cmd
.evt
= C_SET_MULTICAST_MODE
;
2810 cmd
.code
= C_C_MCAST_DISABLE
;
2812 ace_issue_cmd(regs
, &cmd
);
2817 static struct net_device_stats
*ace_get_stats(struct net_device
*dev
)
2819 struct ace_private
*ap
= netdev_priv(dev
);
2820 struct ace_mac_stats __iomem
*mac_stats
=
2821 (struct ace_mac_stats __iomem
*)ap
->regs
->Stats
;
2823 dev
->stats
.rx_missed_errors
= readl(&mac_stats
->drop_space
);
2824 dev
->stats
.multicast
= readl(&mac_stats
->kept_mc
);
2825 dev
->stats
.collisions
= readl(&mac_stats
->coll
);
2831 static void __devinit
ace_copy(struct ace_regs __iomem
*regs
, const __be32
*src
,
2834 void __iomem
*tdest
;
2841 tsize
= min_t(u32
, ((~dest
& (ACE_WINDOW_SIZE
- 1)) + 1),
2842 min_t(u32
, size
, ACE_WINDOW_SIZE
));
2843 tdest
= (void __iomem
*) ®s
->Window
+
2844 (dest
& (ACE_WINDOW_SIZE
- 1));
2845 writel(dest
& ~(ACE_WINDOW_SIZE
- 1), ®s
->WinBase
);
2846 for (i
= 0; i
< (tsize
/ 4); i
++) {
2847 /* Firmware is big-endian */
2848 writel(be32_to_cpup(src
), tdest
);
2858 static void __devinit
ace_clear(struct ace_regs __iomem
*regs
, u32 dest
, int size
)
2860 void __iomem
*tdest
;
2867 tsize
= min_t(u32
, ((~dest
& (ACE_WINDOW_SIZE
- 1)) + 1),
2868 min_t(u32
, size
, ACE_WINDOW_SIZE
));
2869 tdest
= (void __iomem
*) ®s
->Window
+
2870 (dest
& (ACE_WINDOW_SIZE
- 1));
2871 writel(dest
& ~(ACE_WINDOW_SIZE
- 1), ®s
->WinBase
);
2873 for (i
= 0; i
< (tsize
/ 4); i
++) {
2874 writel(0, tdest
+ i
*4);
2884 * Download the firmware into the SRAM on the NIC
2886 * This operation requires the NIC to be halted and is performed with
2887 * interrupts disabled and with the spinlock hold.
2889 static int __devinit
ace_load_firmware(struct net_device
*dev
)
2891 const struct firmware
*fw
;
2892 const char *fw_name
= "acenic/tg2.bin";
2893 struct ace_private
*ap
= netdev_priv(dev
);
2894 struct ace_regs __iomem
*regs
= ap
->regs
;
2895 const __be32
*fw_data
;
2899 if (!(readl(®s
->CpuCtrl
) & CPU_HALTED
)) {
2900 printk(KERN_ERR
"%s: trying to download firmware while the "
2901 "CPU is running!\n", ap
->name
);
2905 if (ACE_IS_TIGON_I(ap
))
2906 fw_name
= "acenic/tg1.bin";
2908 ret
= request_firmware(&fw
, fw_name
, &ap
->pdev
->dev
);
2910 printk(KERN_ERR
"%s: Failed to load firmware \"%s\"\n",
2915 fw_data
= (void *)fw
->data
;
2917 /* Firmware blob starts with version numbers, followed by
2918 load and start address. Remainder is the blob to be loaded
2919 contiguously from load address. We don't bother to represent
2920 the BSS/SBSS sections any more, since we were clearing the
2921 whole thing anyway. */
2922 ap
->firmware_major
= fw
->data
[0];
2923 ap
->firmware_minor
= fw
->data
[1];
2924 ap
->firmware_fix
= fw
->data
[2];
2926 ap
->firmware_start
= be32_to_cpu(fw_data
[1]);
2927 if (ap
->firmware_start
< 0x4000 || ap
->firmware_start
>= 0x80000) {
2928 printk(KERN_ERR
"%s: bogus load address %08x in \"%s\"\n",
2929 ap
->name
, ap
->firmware_start
, fw_name
);
2934 load_addr
= be32_to_cpu(fw_data
[2]);
2935 if (load_addr
< 0x4000 || load_addr
>= 0x80000) {
2936 printk(KERN_ERR
"%s: bogus load address %08x in \"%s\"\n",
2937 ap
->name
, load_addr
, fw_name
);
2943 * Do not try to clear more than 512KiB or we end up seeing
2944 * funny things on NICs with only 512KiB SRAM
2946 ace_clear(regs
, 0x2000, 0x80000-0x2000);
2947 ace_copy(regs
, &fw_data
[3], load_addr
, fw
->size
-12);
2949 release_firmware(fw
);
2955 * The eeprom on the AceNIC is an Atmel i2c EEPROM.
2957 * Accessing the EEPROM is `interesting' to say the least - don't read
2958 * this code right after dinner.
2960 * This is all about black magic and bit-banging the device .... I
2961 * wonder in what hospital they have put the guy who designed the i2c
2964 * Oh yes, this is only the beginning!
2966 * Thanks to Stevarino Webinski for helping tracking down the bugs in the
2967 * code i2c readout code by beta testing all my hacks.
2969 static void __devinit
eeprom_start(struct ace_regs __iomem
*regs
)
2973 readl(®s
->LocalCtrl
);
2974 udelay(ACE_SHORT_DELAY
);
2975 local
= readl(®s
->LocalCtrl
);
2976 local
|= EEPROM_DATA_OUT
| EEPROM_WRITE_ENABLE
;
2977 writel(local
, ®s
->LocalCtrl
);
2978 readl(®s
->LocalCtrl
);
2980 udelay(ACE_SHORT_DELAY
);
2981 local
|= EEPROM_CLK_OUT
;
2982 writel(local
, ®s
->LocalCtrl
);
2983 readl(®s
->LocalCtrl
);
2985 udelay(ACE_SHORT_DELAY
);
2986 local
&= ~EEPROM_DATA_OUT
;
2987 writel(local
, ®s
->LocalCtrl
);
2988 readl(®s
->LocalCtrl
);
2990 udelay(ACE_SHORT_DELAY
);
2991 local
&= ~EEPROM_CLK_OUT
;
2992 writel(local
, ®s
->LocalCtrl
);
2993 readl(®s
->LocalCtrl
);
2998 static void __devinit
eeprom_prep(struct ace_regs __iomem
*regs
, u8 magic
)
3003 udelay(ACE_SHORT_DELAY
);
3004 local
= readl(®s
->LocalCtrl
);
3005 local
&= ~EEPROM_DATA_OUT
;
3006 local
|= EEPROM_WRITE_ENABLE
;
3007 writel(local
, ®s
->LocalCtrl
);
3008 readl(®s
->LocalCtrl
);
3011 for (i
= 0; i
< 8; i
++, magic
<<= 1) {
3012 udelay(ACE_SHORT_DELAY
);
3014 local
|= EEPROM_DATA_OUT
;
3016 local
&= ~EEPROM_DATA_OUT
;
3017 writel(local
, ®s
->LocalCtrl
);
3018 readl(®s
->LocalCtrl
);
3021 udelay(ACE_SHORT_DELAY
);
3022 local
|= EEPROM_CLK_OUT
;
3023 writel(local
, ®s
->LocalCtrl
);
3024 readl(®s
->LocalCtrl
);
3026 udelay(ACE_SHORT_DELAY
);
3027 local
&= ~(EEPROM_CLK_OUT
| EEPROM_DATA_OUT
);
3028 writel(local
, ®s
->LocalCtrl
);
3029 readl(®s
->LocalCtrl
);
3035 static int __devinit
eeprom_check_ack(struct ace_regs __iomem
*regs
)
3040 local
= readl(®s
->LocalCtrl
);
3041 local
&= ~EEPROM_WRITE_ENABLE
;
3042 writel(local
, ®s
->LocalCtrl
);
3043 readl(®s
->LocalCtrl
);
3045 udelay(ACE_LONG_DELAY
);
3046 local
|= EEPROM_CLK_OUT
;
3047 writel(local
, ®s
->LocalCtrl
);
3048 readl(®s
->LocalCtrl
);
3050 udelay(ACE_SHORT_DELAY
);
3051 /* sample data in middle of high clk */
3052 state
= (readl(®s
->LocalCtrl
) & EEPROM_DATA_IN
) != 0;
3053 udelay(ACE_SHORT_DELAY
);
3055 writel(readl(®s
->LocalCtrl
) & ~EEPROM_CLK_OUT
, ®s
->LocalCtrl
);
3056 readl(®s
->LocalCtrl
);
3063 static void __devinit
eeprom_stop(struct ace_regs __iomem
*regs
)
3067 udelay(ACE_SHORT_DELAY
);
3068 local
= readl(®s
->LocalCtrl
);
3069 local
|= EEPROM_WRITE_ENABLE
;
3070 writel(local
, ®s
->LocalCtrl
);
3071 readl(®s
->LocalCtrl
);
3073 udelay(ACE_SHORT_DELAY
);
3074 local
&= ~EEPROM_DATA_OUT
;
3075 writel(local
, ®s
->LocalCtrl
);
3076 readl(®s
->LocalCtrl
);
3078 udelay(ACE_SHORT_DELAY
);
3079 local
|= EEPROM_CLK_OUT
;
3080 writel(local
, ®s
->LocalCtrl
);
3081 readl(®s
->LocalCtrl
);
3083 udelay(ACE_SHORT_DELAY
);
3084 local
|= EEPROM_DATA_OUT
;
3085 writel(local
, ®s
->LocalCtrl
);
3086 readl(®s
->LocalCtrl
);
3088 udelay(ACE_LONG_DELAY
);
3089 local
&= ~EEPROM_CLK_OUT
;
3090 writel(local
, ®s
->LocalCtrl
);
3096 * Read a whole byte from the EEPROM.
3098 static int __devinit
read_eeprom_byte(struct net_device
*dev
,
3099 unsigned long offset
)
3101 struct ace_private
*ap
= netdev_priv(dev
);
3102 struct ace_regs __iomem
*regs
= ap
->regs
;
3103 unsigned long flags
;
3109 * Don't take interrupts on this CPU will bit banging
3110 * the %#%#@$ I2C device
3112 local_irq_save(flags
);
3116 eeprom_prep(regs
, EEPROM_WRITE_SELECT
);
3117 if (eeprom_check_ack(regs
)) {
3118 local_irq_restore(flags
);
3119 printk(KERN_ERR
"%s: Unable to sync eeprom\n", ap
->name
);
3121 goto eeprom_read_error
;
3124 eeprom_prep(regs
, (offset
>> 8) & 0xff);
3125 if (eeprom_check_ack(regs
)) {
3126 local_irq_restore(flags
);
3127 printk(KERN_ERR
"%s: Unable to set address byte 0\n",
3130 goto eeprom_read_error
;
3133 eeprom_prep(regs
, offset
& 0xff);
3134 if (eeprom_check_ack(regs
)) {
3135 local_irq_restore(flags
);
3136 printk(KERN_ERR
"%s: Unable to set address byte 1\n",
3139 goto eeprom_read_error
;
3143 eeprom_prep(regs
, EEPROM_READ_SELECT
);
3144 if (eeprom_check_ack(regs
)) {
3145 local_irq_restore(flags
);
3146 printk(KERN_ERR
"%s: Unable to set READ_SELECT\n",
3149 goto eeprom_read_error
;
3152 for (i
= 0; i
< 8; i
++) {
3153 local
= readl(®s
->LocalCtrl
);
3154 local
&= ~EEPROM_WRITE_ENABLE
;
3155 writel(local
, ®s
->LocalCtrl
);
3156 readl(®s
->LocalCtrl
);
3157 udelay(ACE_LONG_DELAY
);
3159 local
|= EEPROM_CLK_OUT
;
3160 writel(local
, ®s
->LocalCtrl
);
3161 readl(®s
->LocalCtrl
);
3163 udelay(ACE_SHORT_DELAY
);
3164 /* sample data mid high clk */
3165 result
= (result
<< 1) |
3166 ((readl(®s
->LocalCtrl
) & EEPROM_DATA_IN
) != 0);
3167 udelay(ACE_SHORT_DELAY
);
3169 local
= readl(®s
->LocalCtrl
);
3170 local
&= ~EEPROM_CLK_OUT
;
3171 writel(local
, ®s
->LocalCtrl
);
3172 readl(®s
->LocalCtrl
);
3173 udelay(ACE_SHORT_DELAY
);
3176 local
|= EEPROM_WRITE_ENABLE
;
3177 writel(local
, ®s
->LocalCtrl
);
3178 readl(®s
->LocalCtrl
);
3180 udelay(ACE_SHORT_DELAY
);
3184 local
|= EEPROM_DATA_OUT
;
3185 writel(local
, ®s
->LocalCtrl
);
3186 readl(®s
->LocalCtrl
);
3188 udelay(ACE_SHORT_DELAY
);
3189 writel(readl(®s
->LocalCtrl
) | EEPROM_CLK_OUT
, ®s
->LocalCtrl
);
3190 readl(®s
->LocalCtrl
);
3191 udelay(ACE_LONG_DELAY
);
3192 writel(readl(®s
->LocalCtrl
) & ~EEPROM_CLK_OUT
, ®s
->LocalCtrl
);
3193 readl(®s
->LocalCtrl
);
3195 udelay(ACE_SHORT_DELAY
);
3198 local_irq_restore(flags
);
3203 printk(KERN_ERR
"%s: Unable to read eeprom byte 0x%02lx\n",