xref: /netbsd-src/sys/dev/pci/if_bge.c (revision 274254cdae52594c1aa480a736aef78313d15c9c)
1 /*	$NetBSD: if_bge.c,v 1.159 2009/03/22 18:54:59 msaitoh Exp $	*/
2 
3 /*
4  * Copyright (c) 2001 Wind River Systems
5  * Copyright (c) 1997, 1998, 1999, 2001
6  *	Bill Paul <wpaul@windriver.com>.  All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *	This product includes software developed by Bill Paul.
19  * 4. Neither the name of the author nor the names of any co-contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY Bill Paul AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL Bill Paul OR THE VOICES IN HIS HEAD
27  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
28  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
29  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
30  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
31  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
32  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
33  * THE POSSIBILITY OF SUCH DAMAGE.
34  *
35  * $FreeBSD: if_bge.c,v 1.13 2002/04/04 06:01:31 wpaul Exp $
36  */
37 
38 /*
39  * Broadcom BCM570x family gigabit ethernet driver for NetBSD.
40  *
41  * NetBSD version by:
42  *
43  *	Frank van der Linden <fvdl@wasabisystems.com>
44  *	Jason Thorpe <thorpej@wasabisystems.com>
45  *	Jonathan Stone <jonathan@dsg.stanford.edu>
46  *
47  * Originally written for FreeBSD by Bill Paul <wpaul@windriver.com>
48  * Senior Engineer, Wind River Systems
49  */
50 
51 /*
52  * The Broadcom BCM5700 is based on technology originally developed by
53  * Alteon Networks as part of the Tigon I and Tigon II gigabit ethernet
54  * MAC chips. The BCM5700, sometimes refered to as the Tigon III, has
55  * two on-board MIPS R4000 CPUs and can have as much as 16MB of external
56  * SSRAM. The BCM5700 supports TCP, UDP and IP checksum offload, jumbo
57  * frames, highly configurable RX filtering, and 16 RX and TX queues
58  * (which, along with RX filter rules, can be used for QOS applications).
59  * Other features, such as TCP segmentation, may be available as part
60  * of value-added firmware updates. Unlike the Tigon I and Tigon II,
61  * firmware images can be stored in hardware and need not be compiled
62  * into the driver.
63  *
64  * The BCM5700 supports the PCI v2.2 and PCI-X v1.0 standards, and will
65  * function in a 32-bit/64-bit 33/66MHz bus, or a 64-bit/133MHz bus.
66  *
67  * The BCM5701 is a single-chip solution incorporating both the BCM5700
68  * MAC and a BCM5401 10/100/1000 PHY. Unlike the BCM5700, the BCM5701
69  * does not support external SSRAM.
70  *
71  * Broadcom also produces a variation of the BCM5700 under the "Altima"
72  * brand name, which is functionally similar but lacks PCI-X support.
73  *
74  * Without external SSRAM, you can only have at most 4 TX rings,
75  * and the use of the mini RX ring is disabled. This seems to imply
76  * that these features are simply not available on the BCM5701. As a
77  * result, this driver does not implement any support for the mini RX
78  * ring.
79  */
80 
81 #include <sys/cdefs.h>
82 __KERNEL_RCSID(0, "$NetBSD: if_bge.c,v 1.159 2009/03/22 18:54:59 msaitoh Exp $");
83 
84 #include "bpfilter.h"
85 #include "vlan.h"
86 #include "rnd.h"
87 
88 #include <sys/param.h>
89 #include <sys/systm.h>
90 #include <sys/callout.h>
91 #include <sys/sockio.h>
92 #include <sys/mbuf.h>
93 #include <sys/malloc.h>
94 #include <sys/kernel.h>
95 #include <sys/device.h>
96 #include <sys/socket.h>
97 #include <sys/sysctl.h>
98 
99 #include <net/if.h>
100 #include <net/if_dl.h>
101 #include <net/if_media.h>
102 #include <net/if_ether.h>
103 
104 #if NRND > 0
105 #include <sys/rnd.h>
106 #endif
107 
108 #ifdef INET
109 #include <netinet/in.h>
110 #include <netinet/in_systm.h>
111 #include <netinet/in_var.h>
112 #include <netinet/ip.h>
113 #endif
114 
115 /* Headers for TCP  Segmentation Offload (TSO) */
116 #include <netinet/in_systm.h>		/* n_time for <netinet/ip.h>... */
117 #include <netinet/in.h>			/* ip_{src,dst}, for <netinet/ip.h> */
118 #include <netinet/ip.h>			/* for struct ip */
119 #include <netinet/tcp.h>		/* for struct tcphdr */
120 
121 
122 #if NBPFILTER > 0
123 #include <net/bpf.h>
124 #endif
125 
126 #include <dev/pci/pcireg.h>
127 #include <dev/pci/pcivar.h>
128 #include <dev/pci/pcidevs.h>
129 
130 #include <dev/mii/mii.h>
131 #include <dev/mii/miivar.h>
132 #include <dev/mii/miidevs.h>
133 #include <dev/mii/brgphyreg.h>
134 
135 #include <dev/pci/if_bgereg.h>
136 
137 #include <uvm/uvm_extern.h>
138 
139 #define ETHER_MIN_NOPAD (ETHER_MIN_LEN - ETHER_CRC_LEN) /* i.e., 60 */
140 
141 
142 /*
143  * Tunable thresholds for rx-side bge interrupt mitigation.
144  */
145 
146 /*
147  * The pairs of values below were obtained from empirical measurement
148  * on bcm5700 rev B2; they ar designed to give roughly 1 receive
149  * interrupt for every N packets received, where N is, approximately,
150  * the second value (rx_max_bds) in each pair.  The values are chosen
151  * such that moving from one pair to the succeeding pair was observed
152  * to roughly halve interrupt rate under sustained input packet load.
153  * The values were empirically chosen to avoid overflowing internal
154  * limits on the  bcm5700: inreasing rx_ticks much beyond 600
155  * results in internal wrapping and higher interrupt rates.
156  * The limit of 46 frames was chosen to match NFS workloads.
157  *
158  * These values also work well on bcm5701, bcm5704C, and (less
159  * tested) bcm5703.  On other chipsets, (including the Altima chip
160  * family), the larger values may overflow internal chip limits,
161  * leading to increasing interrupt rates rather than lower interrupt
162  * rates.
163  *
164  * Applications using heavy interrupt mitigation (interrupting every
165  * 32 or 46 frames) in both directions may need to increase the TCP
166  * windowsize to above 131072 bytes (e.g., to 199608 bytes) to sustain
167  * full link bandwidth, due to ACKs and window updates lingering
168  * in the RX queue during the 30-to-40-frame interrupt-mitigation window.
169  */
170 static const struct bge_load_rx_thresh {
171 	int rx_ticks;
172 	int rx_max_bds; }
173 bge_rx_threshes[] = {
174 	{ 32,   2 },
175 	{ 50,   4 },
176 	{ 100,  8 },
177 	{ 192, 16 },
178 	{ 416, 32 },
179 	{ 598, 46 }
180 };
181 #define NBGE_RX_THRESH (sizeof(bge_rx_threshes) / sizeof(bge_rx_threshes[0]))
182 
183 /* XXX patchable; should be sysctl'able */
184 static int	bge_auto_thresh = 1;
185 static int	bge_rx_thresh_lvl;
186 
187 static int	bge_rxthresh_nodenum;
188 
189 typedef int (*bge_eaddr_fcn_t)(struct bge_softc *, u_int8_t[]);
190 
191 static int	bge_probe(device_t, cfdata_t, void *);
192 static void	bge_attach(device_t, device_t, void *);
193 static void	bge_release_resources(struct bge_softc *);
194 static void	bge_txeof(struct bge_softc *);
195 static void	bge_rxeof(struct bge_softc *);
196 
197 static int 	bge_get_eaddr_mem(struct bge_softc *, u_int8_t[]);
198 static int 	bge_get_eaddr_nvram(struct bge_softc *, u_int8_t[]);
199 static int 	bge_get_eaddr_eeprom(struct bge_softc *, u_int8_t[]);
200 static int 	bge_get_eaddr(struct bge_softc *, u_int8_t[]);
201 
202 static void	bge_tick(void *);
203 static void	bge_stats_update(struct bge_softc *);
204 static int	bge_encap(struct bge_softc *, struct mbuf *, u_int32_t *);
205 
206 static int	bge_intr(void *);
207 static void	bge_start(struct ifnet *);
208 static int	bge_ioctl(struct ifnet *, u_long, void *);
209 static int	bge_init(struct ifnet *);
210 static void	bge_stop(struct ifnet *, int);
211 static void	bge_watchdog(struct ifnet *);
212 static int	bge_ifmedia_upd(struct ifnet *);
213 static void	bge_ifmedia_sts(struct ifnet *, struct ifmediareq *);
214 
215 static void	bge_setmulti(struct bge_softc *);
216 
217 static void	bge_handle_events(struct bge_softc *);
218 static int	bge_alloc_jumbo_mem(struct bge_softc *);
219 #if 0 /* XXX */
220 static void	bge_free_jumbo_mem(struct bge_softc *);
221 #endif
222 static void	*bge_jalloc(struct bge_softc *);
223 static void	bge_jfree(struct mbuf *, void *, size_t, void *);
224 static int	bge_newbuf_std(struct bge_softc *, int, struct mbuf *,
225 			       bus_dmamap_t);
226 static int	bge_newbuf_jumbo(struct bge_softc *, int, struct mbuf *);
227 static int	bge_init_rx_ring_std(struct bge_softc *);
228 static void	bge_free_rx_ring_std(struct bge_softc *);
229 static int	bge_init_rx_ring_jumbo(struct bge_softc *);
230 static void	bge_free_rx_ring_jumbo(struct bge_softc *);
231 static void	bge_free_tx_ring(struct bge_softc *);
232 static int	bge_init_tx_ring(struct bge_softc *);
233 
234 static int	bge_chipinit(struct bge_softc *);
235 static int	bge_blockinit(struct bge_softc *);
236 static int	bge_setpowerstate(struct bge_softc *, int);
237 
238 static void	bge_reset(struct bge_softc *);
239 
240 #define BGE_DEBUG
241 #ifdef BGE_DEBUG
242 #define DPRINTF(x)	if (bgedebug) printf x
243 #define DPRINTFN(n,x)	if (bgedebug >= (n)) printf x
244 #define BGE_TSO_PRINTF(x)  do { if (bge_tso_debug) printf x ;} while (0)
245 int	bgedebug = 0;
246 int	bge_tso_debug = 0;
247 #else
248 #define DPRINTF(x)
249 #define DPRINTFN(n,x)
250 #define BGE_TSO_PRINTF(x)
251 #endif
252 
253 #ifdef BGE_EVENT_COUNTERS
254 #define	BGE_EVCNT_INCR(ev)	(ev).ev_count++
255 #define	BGE_EVCNT_ADD(ev, val)	(ev).ev_count += (val)
256 #define	BGE_EVCNT_UPD(ev, val)	(ev).ev_count = (val)
257 #else
258 #define	BGE_EVCNT_INCR(ev)	/* nothing */
259 #define	BGE_EVCNT_ADD(ev, val)	/* nothing */
260 #define	BGE_EVCNT_UPD(ev, val)	/* nothing */
261 #endif
262 
263 static const struct bge_product {
264 	pci_vendor_id_t		bp_vendor;
265 	pci_product_id_t	bp_product;
266 	const char		*bp_name;
267 } bge_products[] = {
268 	/*
269 	 * The BCM5700 documentation seems to indicate that the hardware
270 	 * still has the Alteon vendor ID burned into it, though it
271 	 * should always be overridden by the value in the EEPROM.  We'll
272 	 * check for it anyway.
273 	 */
274 	{ PCI_VENDOR_ALTEON,
275 	  PCI_PRODUCT_ALTEON_BCM5700,
276 	  "Broadcom BCM5700 Gigabit Ethernet",
277 	  },
278 	{ PCI_VENDOR_ALTEON,
279 	  PCI_PRODUCT_ALTEON_BCM5701,
280 	  "Broadcom BCM5701 Gigabit Ethernet",
281 	  },
282 	{ PCI_VENDOR_ALTIMA,
283 	  PCI_PRODUCT_ALTIMA_AC1000,
284 	  "Altima AC1000 Gigabit Ethernet",
285 	  },
286 	{ PCI_VENDOR_ALTIMA,
287 	  PCI_PRODUCT_ALTIMA_AC1001,
288 	  "Altima AC1001 Gigabit Ethernet",
289 	   },
290 	{ PCI_VENDOR_ALTIMA,
291 	  PCI_PRODUCT_ALTIMA_AC9100,
292 	  "Altima AC9100 Gigabit Ethernet",
293 	  },
294 	{ PCI_VENDOR_BROADCOM,
295 	  PCI_PRODUCT_BROADCOM_BCM5700,
296 	  "Broadcom BCM5700 Gigabit Ethernet",
297 	  },
298 	{ PCI_VENDOR_BROADCOM,
299 	  PCI_PRODUCT_BROADCOM_BCM5701,
300 	  "Broadcom BCM5701 Gigabit Ethernet",
301 	  },
302 	{ PCI_VENDOR_BROADCOM,
303 	  PCI_PRODUCT_BROADCOM_BCM5702,
304 	  "Broadcom BCM5702 Gigabit Ethernet",
305 	  },
306 	{ PCI_VENDOR_BROADCOM,
307 	  PCI_PRODUCT_BROADCOM_BCM5702X,
308 	  "Broadcom BCM5702X Gigabit Ethernet" },
309 	{ PCI_VENDOR_BROADCOM,
310 	  PCI_PRODUCT_BROADCOM_BCM5703,
311 	  "Broadcom BCM5703 Gigabit Ethernet",
312 	  },
313 	{ PCI_VENDOR_BROADCOM,
314 	  PCI_PRODUCT_BROADCOM_BCM5703X,
315 	  "Broadcom BCM5703X Gigabit Ethernet",
316 	  },
317 	{ PCI_VENDOR_BROADCOM,
318 	  PCI_PRODUCT_BROADCOM_BCM5703_ALT,
319 	  "Broadcom BCM5703 Gigabit Ethernet",
320 	  },
321    	{ PCI_VENDOR_BROADCOM,
322 	  PCI_PRODUCT_BROADCOM_BCM5704C,
323 	  "Broadcom BCM5704C Dual Gigabit Ethernet",
324 	  },
325    	{ PCI_VENDOR_BROADCOM,
326 	  PCI_PRODUCT_BROADCOM_BCM5704S,
327 	  "Broadcom BCM5704S Dual Gigabit Ethernet",
328 	  },
329    	{ PCI_VENDOR_BROADCOM,
330 	  PCI_PRODUCT_BROADCOM_BCM5705,
331 	  "Broadcom BCM5705 Gigabit Ethernet",
332 	  },
333    	{ PCI_VENDOR_BROADCOM,
334 	  PCI_PRODUCT_BROADCOM_BCM5705K,
335 	  "Broadcom BCM5705K Gigabit Ethernet",
336 	  },
337    	{ PCI_VENDOR_BROADCOM,
338 	  PCI_PRODUCT_BROADCOM_BCM5705M,
339 	  "Broadcom BCM5705M Gigabit Ethernet",
340 	  },
341    	{ PCI_VENDOR_BROADCOM,
342 	  PCI_PRODUCT_BROADCOM_BCM5705M_ALT,
343 	  "Broadcom BCM5705M Gigabit Ethernet",
344 	  },
345 	{ PCI_VENDOR_BROADCOM,
346 	  PCI_PRODUCT_BROADCOM_BCM5714,
347 	  "Broadcom BCM5714/5715 Gigabit Ethernet",
348 	  },
349 	{ PCI_VENDOR_BROADCOM,
350 	  PCI_PRODUCT_BROADCOM_BCM5715,
351 	  "Broadcom BCM5714/5715 Gigabit Ethernet",
352 	  },
353 	{ PCI_VENDOR_BROADCOM,
354 	  PCI_PRODUCT_BROADCOM_BCM5789,
355 	  "Broadcom BCM5789 Gigabit Ethernet",
356 	  },
357 	{ PCI_VENDOR_BROADCOM,
358 	  PCI_PRODUCT_BROADCOM_BCM5721,
359 	  "Broadcom BCM5721 Gigabit Ethernet",
360 	  },
361 	{ PCI_VENDOR_BROADCOM,
362 	  PCI_PRODUCT_BROADCOM_BCM5722,
363 	  "Broadcom BCM5722 Gigabit Ethernet",
364 	  },
365 	{ PCI_VENDOR_BROADCOM,
366 	  PCI_PRODUCT_BROADCOM_BCM5750,
367 	  "Broadcom BCM5750 Gigabit Ethernet",
368 	  },
369 	{ PCI_VENDOR_BROADCOM,
370 	  PCI_PRODUCT_BROADCOM_BCM5750M,
371 	  "Broadcom BCM5750M Gigabit Ethernet",
372 	  },
373 	{ PCI_VENDOR_BROADCOM,
374 	  PCI_PRODUCT_BROADCOM_BCM5751,
375 	  "Broadcom BCM5751 Gigabit Ethernet",
376 	  },
377 	{ PCI_VENDOR_BROADCOM,
378 	  PCI_PRODUCT_BROADCOM_BCM5751M,
379 	  "Broadcom BCM5751M Gigabit Ethernet",
380 	  },
381 	{ PCI_VENDOR_BROADCOM,
382 	  PCI_PRODUCT_BROADCOM_BCM5752,
383 	  "Broadcom BCM5752 Gigabit Ethernet",
384 	  },
385 	{ PCI_VENDOR_BROADCOM,
386 	  PCI_PRODUCT_BROADCOM_BCM5752M,
387 	  "Broadcom BCM5752M Gigabit Ethernet",
388 	  },
389 	{ PCI_VENDOR_BROADCOM,
390 	  PCI_PRODUCT_BROADCOM_BCM5753,
391 	  "Broadcom BCM5753 Gigabit Ethernet",
392 	  },
393 	{ PCI_VENDOR_BROADCOM,
394 	  PCI_PRODUCT_BROADCOM_BCM5753M,
395 	  "Broadcom BCM5753M Gigabit Ethernet",
396 	  },
397 	{ PCI_VENDOR_BROADCOM,
398 	  PCI_PRODUCT_BROADCOM_BCM5754,
399 	  "Broadcom BCM5754 Gigabit Ethernet",
400 	},
401 	{ PCI_VENDOR_BROADCOM,
402 	  PCI_PRODUCT_BROADCOM_BCM5754M,
403 	  "Broadcom BCM5754M Gigabit Ethernet",
404 	},
405 	{ PCI_VENDOR_BROADCOM,
406 	  PCI_PRODUCT_BROADCOM_BCM5755,
407 	  "Broadcom BCM5755 Gigabit Ethernet",
408 	},
409 	{ PCI_VENDOR_BROADCOM,
410 	  PCI_PRODUCT_BROADCOM_BCM5755M,
411 	  "Broadcom BCM5755M Gigabit Ethernet",
412 	},
413    	{ PCI_VENDOR_BROADCOM,
414 	  PCI_PRODUCT_BROADCOM_BCM5780,
415 	  "Broadcom BCM5780 Gigabit Ethernet",
416 	  },
417    	{ PCI_VENDOR_BROADCOM,
418 	  PCI_PRODUCT_BROADCOM_BCM5780S,
419 	  "Broadcom BCM5780S Gigabit Ethernet",
420 	  },
421    	{ PCI_VENDOR_BROADCOM,
422 	  PCI_PRODUCT_BROADCOM_BCM5782,
423 	  "Broadcom BCM5782 Gigabit Ethernet",
424 	},
425 	{ PCI_VENDOR_BROADCOM,
426 	  PCI_PRODUCT_BROADCOM_BCM5786,
427 	  "Broadcom BCM5786 Gigabit Ethernet",
428 	},
429 	{ PCI_VENDOR_BROADCOM,
430 	  PCI_PRODUCT_BROADCOM_BCM5787,
431 	  "Broadcom BCM5787 Gigabit Ethernet",
432 	},
433 	{ PCI_VENDOR_BROADCOM,
434 	  PCI_PRODUCT_BROADCOM_BCM5787M,
435 	  "Broadcom BCM5787M Gigabit Ethernet",
436 	},
437    	{ PCI_VENDOR_BROADCOM,
438 	  PCI_PRODUCT_BROADCOM_BCM5788,
439 	  "Broadcom BCM5788 Gigabit Ethernet",
440 	  },
441    	{ PCI_VENDOR_BROADCOM,
442 	  PCI_PRODUCT_BROADCOM_BCM5789,
443 	  "Broadcom BCM5789 Gigabit Ethernet",
444 	  },
445    	{ PCI_VENDOR_BROADCOM,
446 	  PCI_PRODUCT_BROADCOM_BCM5901,
447 	  "Broadcom BCM5901 Fast Ethernet",
448 	  },
449    	{ PCI_VENDOR_BROADCOM,
450 	  PCI_PRODUCT_BROADCOM_BCM5901A2,
451 	  "Broadcom BCM5901A2 Fast Ethernet",
452 	  },
453 	{ PCI_VENDOR_SCHNEIDERKOCH,
454 	  PCI_PRODUCT_SCHNEIDERKOCH_SK_9DX1,
455 	  "SysKonnect SK-9Dx1 Gigabit Ethernet",
456 	  },
457 	{ PCI_VENDOR_3COM,
458 	  PCI_PRODUCT_3COM_3C996,
459 	  "3Com 3c996 Gigabit Ethernet",
460 	  },
461 	{ PCI_VENDOR_BROADCOM,
462 	  PCI_PRODUCT_BROADCOM_BCM5906,
463 	  "Broadcom BCM5906 Fast Ethernet",
464 	  },
465 	{ PCI_VENDOR_BROADCOM,
466 	  PCI_PRODUCT_BROADCOM_BCM5906M,
467 	  "Broadcom BCM5906M Fast Ethernet",
468 	  },
469 	{ 0,
470 	  0,
471 	  NULL },
472 };
473 
474 /*
475  * XXX: how to handle variants based on 5750 and derivatives:
476  * 5750 5751, 5721, possibly 5714, 5752, and 5708?, which
477  * in general behave like a 5705, except with additional quirks.
478  * This driver's current handling of the 5721 is wrong;
479  * how we map ASIC revision to "quirks" needs more thought.
480  * (defined here until the thought is done).
481  */
482 #define BGE_IS_5714_FAMILY(sc) \
483 	(BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5714_A0 || \
484 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5780 ||	\
485 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5714 )
486 
487 #define BGE_IS_5750_OR_BEYOND(sc)  \
488 	(BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5750 || \
489 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5752 || \
490 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5755 || \
491 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5787 || \
492 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5906 || \
493 	 BGE_IS_5714_FAMILY(sc) )
494 
495 #define BGE_IS_5705_OR_BEYOND(sc)  \
496 	(BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5705 || \
497 	    (BGE_IS_5750_OR_BEYOND(sc)))
498 
499 static const struct bge_revision {
500 	uint32_t		br_chipid;
501 	const char		*br_name;
502 } bge_revisions[] = {
503 	{ BGE_CHIPID_BCM5700_A0, "BCM5700 A0" },
504 	{ BGE_CHIPID_BCM5700_A1, "BCM5700 A1" },
505 	{ BGE_CHIPID_BCM5700_B0, "BCM5700 B0" },
506 	{ BGE_CHIPID_BCM5700_B1, "BCM5700 B1" },
507 	{ BGE_CHIPID_BCM5700_B2, "BCM5700 B2" },
508 	{ BGE_CHIPID_BCM5700_B3, "BCM5700 B3" },
509 	/* This is treated like a BCM5700 Bx */
510 	{ BGE_CHIPID_BCM5700_ALTIMA, "BCM5700 Altima" },
511 	{ BGE_CHIPID_BCM5700_C0, "BCM5700 C0" },
512 	{ BGE_CHIPID_BCM5701_A0, "BCM5701 A0" },
513 	{ BGE_CHIPID_BCM5701_B0, "BCM5701 B0" },
514 	{ BGE_CHIPID_BCM5701_B2, "BCM5701 B2" },
515 	{ BGE_CHIPID_BCM5701_B5, "BCM5701 B5" },
516 	{ BGE_CHIPID_BCM5703_A0, "BCM5703 A0" },
517 	{ BGE_CHIPID_BCM5703_A1, "BCM5703 A1" },
518 	{ BGE_CHIPID_BCM5703_A2, "BCM5703 A2" },
519 	{ BGE_CHIPID_BCM5703_A3, "BCM5703 A3" },
520 	{ BGE_CHIPID_BCM5703_B0, "BCM5703 B0" },
521 	{ BGE_CHIPID_BCM5704_A0, "BCM5704 A0" },
522 	{ BGE_CHIPID_BCM5704_A1, "BCM5704 A1" },
523 	{ BGE_CHIPID_BCM5704_A2, "BCM5704 A2" },
524 	{ BGE_CHIPID_BCM5704_A3, "BCM5704 A3" },
525 	{ BGE_CHIPID_BCM5704_B0, "BCM5704 B0" },
526 	{ BGE_CHIPID_BCM5705_A0, "BCM5705 A0" },
527 	{ BGE_CHIPID_BCM5705_A1, "BCM5705 A1" },
528 	{ BGE_CHIPID_BCM5705_A2, "BCM5705 A2" },
529 	{ BGE_CHIPID_BCM5705_A3, "BCM5705 A3" },
530 	{ BGE_CHIPID_BCM5750_A0, "BCM5750 A0" },
531 	{ BGE_CHIPID_BCM5750_A1, "BCM5750 A1" },
532 	{ BGE_CHIPID_BCM5751_A1, "BCM5751 A1" },
533 	{ BGE_CHIPID_BCM5752_A0, "BCM5752 A0" },
534 	{ BGE_CHIPID_BCM5752_A1, "BCM5752 A1" },
535 	{ BGE_CHIPID_BCM5752_A2, "BCM5752 A2" },
536 	{ BGE_CHIPID_BCM5714_A0, "BCM5714 A0" },
537 	{ BGE_CHIPID_BCM5714_B0, "BCM5714 B0" },
538 	{ BGE_CHIPID_BCM5714_B3, "BCM5714 B3" },
539 	{ BGE_CHIPID_BCM5715_A0, "BCM5715 A0" },
540 	{ BGE_CHIPID_BCM5715_A1, "BCM5715 A1" },
541 	{ BGE_CHIPID_BCM5715_A3, "BCM5715 A3" },
542 	{ BGE_CHIPID_BCM5755_A0, "BCM5755 A0" },
543 	{ BGE_CHIPID_BCM5755_A1, "BCM5755 A1" },
544 	{ BGE_CHIPID_BCM5755_A2, "BCM5755 A2" },
545 	{ BGE_CHIPID_BCM5755_C0, "BCM5755 C0" },
546 	{ BGE_CHIPID_BCM5787_A0, "BCM5754/5787 A0" },
547 	{ BGE_CHIPID_BCM5787_A1, "BCM5754/5787 A1" },
548 	{ BGE_CHIPID_BCM5787_A2, "BCM5754/5787 A2" },
549 	{ 0, NULL }
550 };
551 
552 /*
553  * Some defaults for major revisions, so that newer steppings
554  * that we don't know about have a shot at working.
555  */
556 static const struct bge_revision bge_majorrevs[] = {
557 	{ BGE_ASICREV_BCM5700, "unknown BCM5700" },
558 	{ BGE_ASICREV_BCM5701, "unknown BCM5701" },
559 	{ BGE_ASICREV_BCM5703, "unknown BCM5703" },
560 	{ BGE_ASICREV_BCM5704, "unknown BCM5704" },
561 	{ BGE_ASICREV_BCM5705, "unknown BCM5705" },
562 	{ BGE_ASICREV_BCM5750, "unknown BCM575x family" },
563 	{ BGE_ASICREV_BCM5714_A0, "unknown BCM5714" },
564 	{ BGE_ASICREV_BCM5714, "unknown BCM5714" },
565 	{ BGE_ASICREV_BCM5752, "unknown BCM5752 family" },
566 	{ BGE_ASICREV_BCM5755, "unknown BCM5755" },
567 	{ BGE_ASICREV_BCM5780, "unknown BCM5780" },
568 	{ BGE_ASICREV_BCM5787, "unknown BCM5787" },
569 	{ BGE_ASICREV_BCM5906, "unknown BCM5906" },
570 	{ 0, NULL }
571 };
572 
573 CFATTACH_DECL_NEW(bge, sizeof(struct bge_softc),
574     bge_probe, bge_attach, NULL, NULL);
575 
576 static u_int32_t
577 bge_readmem_ind(struct bge_softc *sc, int off)
578 {
579 	pcireg_t val;
580 
581 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MEMWIN_BASEADDR, off);
582 	val = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MEMWIN_DATA);
583 	return val;
584 }
585 
586 static void
587 bge_writemem_ind(struct bge_softc *sc, int off, int val)
588 {
589 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MEMWIN_BASEADDR, off);
590 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MEMWIN_DATA, val);
591 }
592 
593 #ifdef notdef
594 static u_int32_t
595 bge_readreg_ind(struct bge_softc *sc, int off)
596 {
597 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_REG_BASEADDR, off);
598 	return (pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_REG_DATA));
599 }
600 #endif
601 
602 static void
603 bge_writereg_ind(struct bge_softc *sc, int off, int val)
604 {
605 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_REG_BASEADDR, off);
606 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_REG_DATA, val);
607 }
608 
609 static void
610 bge_writemem_direct(struct bge_softc *sc, int off, int val)
611 {
612 	CSR_WRITE_4(sc, off, val);
613 }
614 
615 static void
616 bge_writembx(struct bge_softc *sc, int off, int val)
617 {
618 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5906)
619 		off += BGE_LPMBX_IRQ0_HI - BGE_MBX_IRQ0_HI;
620 
621 	CSR_WRITE_4(sc, off, val);
622 }
623 
624 static u_int8_t
625 bge_nvram_getbyte(struct bge_softc *sc, int addr, u_int8_t *dest)
626 {
627 	u_int32_t access, byte = 0;
628 	int i;
629 
630 	/* Lock. */
631 	CSR_WRITE_4(sc, BGE_NVRAM_SWARB, BGE_NVRAMSWARB_SET1);
632 	for (i = 0; i < 8000; i++) {
633 		if (CSR_READ_4(sc, BGE_NVRAM_SWARB) & BGE_NVRAMSWARB_GNT1)
634 			break;
635 		DELAY(20);
636 	}
637 	if (i == 8000)
638 		return (1);
639 
640 	/* Enable access. */
641 	access = CSR_READ_4(sc, BGE_NVRAM_ACCESS);
642 	CSR_WRITE_4(sc, BGE_NVRAM_ACCESS, access | BGE_NVRAMACC_ENABLE);
643 
644 	CSR_WRITE_4(sc, BGE_NVRAM_ADDR, addr & 0xfffffffc);
645 	CSR_WRITE_4(sc, BGE_NVRAM_CMD, BGE_NVRAM_READCMD);
646 	for (i = 0; i < BGE_TIMEOUT * 10; i++) {
647 		DELAY(10);
648 		if (CSR_READ_4(sc, BGE_NVRAM_CMD) & BGE_NVRAMCMD_DONE) {
649 			DELAY(10);
650 			break;
651 		}
652 	}
653 
654 	if (i == BGE_TIMEOUT * 10) {
655 		aprint_error_dev(sc->bge_dev, "nvram read timed out\n");
656 		return (1);
657 	}
658 
659 	/* Get result. */
660 	byte = CSR_READ_4(sc, BGE_NVRAM_RDDATA);
661 
662 	*dest = (bswap32(byte) >> ((addr % 4) * 8)) & 0xFF;
663 
664 	/* Disable access. */
665 	CSR_WRITE_4(sc, BGE_NVRAM_ACCESS, access);
666 
667 	/* Unlock. */
668 	CSR_WRITE_4(sc, BGE_NVRAM_SWARB, BGE_NVRAMSWARB_CLR1);
669 	CSR_READ_4(sc, BGE_NVRAM_SWARB);
670 
671 	return (0);
672 }
673 
674 /*
675  * Read a sequence of bytes from NVRAM.
676  */
677 static int
678 bge_read_nvram(struct bge_softc *sc, u_int8_t *dest, int off, int cnt)
679 {
680 	int err = 0, i;
681 	u_int8_t byte = 0;
682 
683 	if (BGE_ASICREV(sc->bge_chipid) != BGE_ASICREV_BCM5906)
684 		return (1);
685 
686 	for (i = 0; i < cnt; i++) {
687 		err = bge_nvram_getbyte(sc, off + i, &byte);
688 		if (err)
689 			break;
690 		*(dest + i) = byte;
691 	}
692 
693 	return (err ? 1 : 0);
694 }
695 
696 /*
697  * Read a byte of data stored in the EEPROM at address 'addr.' The
698  * BCM570x supports both the traditional bitbang interface and an
699  * auto access interface for reading the EEPROM. We use the auto
700  * access method.
701  */
702 static u_int8_t
703 bge_eeprom_getbyte(struct bge_softc *sc, int addr, u_int8_t *dest)
704 {
705 	int i;
706 	u_int32_t byte = 0;
707 
708 	/*
709 	 * Enable use of auto EEPROM access so we can avoid
710 	 * having to use the bitbang method.
711 	 */
712 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_AUTO_EEPROM);
713 
714 	/* Reset the EEPROM, load the clock period. */
715 	CSR_WRITE_4(sc, BGE_EE_ADDR,
716 	    BGE_EEADDR_RESET|BGE_EEHALFCLK(BGE_HALFCLK_384SCL));
717 	DELAY(20);
718 
719 	/* Issue the read EEPROM command. */
720 	CSR_WRITE_4(sc, BGE_EE_ADDR, BGE_EE_READCMD | addr);
721 
722 	/* Wait for completion */
723 	for(i = 0; i < BGE_TIMEOUT * 10; i++) {
724 		DELAY(10);
725 		if (CSR_READ_4(sc, BGE_EE_ADDR) & BGE_EEADDR_DONE)
726 			break;
727 	}
728 
729 	if (i == BGE_TIMEOUT) {
730 		aprint_error_dev(sc->bge_dev, "eeprom read timed out\n");
731 		return (0);
732 	}
733 
734 	/* Get result. */
735 	byte = CSR_READ_4(sc, BGE_EE_DATA);
736 
737 	*dest = (byte >> ((addr % 4) * 8)) & 0xFF;
738 
739 	return (0);
740 }
741 
742 /*
743  * Read a sequence of bytes from the EEPROM.
744  */
745 static int
746 bge_read_eeprom(struct bge_softc *sc, void *destv, int off, int cnt)
747 {
748 	int err = 0, i;
749 	u_int8_t byte = 0;
750 	char *dest = destv;
751 
752 	for (i = 0; i < cnt; i++) {
753 		err = bge_eeprom_getbyte(sc, off + i, &byte);
754 		if (err)
755 			break;
756 		*(dest + i) = byte;
757 	}
758 
759 	return (err ? 1 : 0);
760 }
761 
762 static int
763 bge_miibus_readreg(device_t dev, int phy, int reg)
764 {
765 	struct bge_softc *sc = device_private(dev);
766 	u_int32_t val;
767 	u_int32_t saved_autopoll;
768 	int i;
769 
770 	/*
771 	 * Broadcom's own driver always assumes the internal
772 	 * PHY is at GMII address 1. On some chips, the PHY responds
773 	 * to accesses at all addresses, which could cause us to
774 	 * bogusly attach the PHY 32 times at probe type. Always
775 	 * restricting the lookup to address 1 is simpler than
776 	 * trying to figure out which chips revisions should be
777 	 * special-cased.
778 	 */
779 	if (phy != 1)
780 		return (0);
781 
782 	/* Reading with autopolling on may trigger PCI errors */
783 	saved_autopoll = CSR_READ_4(sc, BGE_MI_MODE);
784 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
785 		CSR_WRITE_4(sc, BGE_MI_MODE,
786 		    saved_autopoll &~ BGE_MIMODE_AUTOPOLL);
787 		DELAY(40);
788 	}
789 
790 	CSR_WRITE_4(sc, BGE_MI_COMM, BGE_MICMD_READ|BGE_MICOMM_BUSY|
791 	    BGE_MIPHY(phy)|BGE_MIREG(reg));
792 
793 	for (i = 0; i < BGE_TIMEOUT; i++) {
794 		val = CSR_READ_4(sc, BGE_MI_COMM);
795 		if (!(val & BGE_MICOMM_BUSY))
796 			break;
797 		delay(10);
798 	}
799 
800 	if (i == BGE_TIMEOUT) {
801 		aprint_error_dev(sc->bge_dev, "PHY read timed out\n");
802 		val = 0;
803 		goto done;
804 	}
805 
806 	val = CSR_READ_4(sc, BGE_MI_COMM);
807 
808 done:
809 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
810 		CSR_WRITE_4(sc, BGE_MI_MODE, saved_autopoll);
811 		DELAY(40);
812 	}
813 
814 	if (val & BGE_MICOMM_READFAIL)
815 		return (0);
816 
817 	return (val & 0xFFFF);
818 }
819 
820 static void
821 bge_miibus_writereg(device_t dev, int phy, int reg, int val)
822 {
823 	struct bge_softc *sc = device_private(dev);
824 	u_int32_t saved_autopoll;
825 	int i;
826 
827 	if (phy!=1) {
828 		return;
829 	}
830 
831 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5906 &&
832 	    (reg == BRGPHY_MII_1000CTL || reg == BRGPHY_MII_AUXCTL)) {
833 		return;
834 	}
835 
836 	/* Touching the PHY while autopolling is on may trigger PCI errors */
837 	saved_autopoll = CSR_READ_4(sc, BGE_MI_MODE);
838 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
839 		delay(40);
840 		CSR_WRITE_4(sc, BGE_MI_MODE,
841 		    saved_autopoll & (~BGE_MIMODE_AUTOPOLL));
842 		delay(10); /* 40 usec is supposed to be adequate */
843 	}
844 
845 	CSR_WRITE_4(sc, BGE_MI_COMM, BGE_MICMD_WRITE|BGE_MICOMM_BUSY|
846 	    BGE_MIPHY(phy)|BGE_MIREG(reg)|val);
847 
848 	for (i = 0; i < BGE_TIMEOUT; i++) {
849 		delay(10);
850 		if (!(CSR_READ_4(sc, BGE_MI_COMM) & BGE_MICOMM_BUSY)) {
851 			delay(5);
852 			CSR_READ_4(sc, BGE_MI_COMM);
853 			break;
854 		}
855 	}
856 
857 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
858 		CSR_WRITE_4(sc, BGE_MI_MODE, saved_autopoll);
859 		delay(40);
860 	}
861 
862 	if (i == BGE_TIMEOUT)
863 		aprint_error_dev(sc->bge_dev, "PHY read timed out\n");
864 }
865 
866 static void
867 bge_miibus_statchg(device_t dev)
868 {
869 	struct bge_softc *sc = device_private(dev);
870 	struct mii_data *mii = &sc->bge_mii;
871 
872 	/*
873 	 * Get flow control negotiation result.
874 	 */
875 	if (IFM_SUBTYPE(mii->mii_media.ifm_cur->ifm_media) == IFM_AUTO &&
876 	    (mii->mii_media_active & IFM_ETH_FMASK) != sc->bge_flowflags) {
877 		sc->bge_flowflags = mii->mii_media_active & IFM_ETH_FMASK;
878 		mii->mii_media_active &= ~IFM_ETH_FMASK;
879 	}
880 
881 	BGE_CLRBIT(sc, BGE_MAC_MODE, BGE_MACMODE_PORTMODE);
882 	if (IFM_SUBTYPE(mii->mii_media_active) == IFM_1000_T) {
883 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_PORTMODE_GMII);
884 	} else
885 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_PORTMODE_MII);
886 
887 	if ((mii->mii_media_active & IFM_GMASK) == IFM_FDX)
888 		BGE_CLRBIT(sc, BGE_MAC_MODE, BGE_MACMODE_HALF_DUPLEX);
889 	else
890 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_MACMODE_HALF_DUPLEX);
891 
892 	/*
893 	 * 802.3x flow control
894 	 */
895 	if (sc->bge_flowflags & IFM_ETH_RXPAUSE)
896 		BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_FLOWCTL_ENABLE);
897 	else
898 		BGE_CLRBIT(sc, BGE_RX_MODE, BGE_RXMODE_FLOWCTL_ENABLE);
899 
900 	if (sc->bge_flowflags & IFM_ETH_TXPAUSE)
901 		BGE_SETBIT(sc, BGE_TX_MODE, BGE_TXMODE_FLOWCTL_ENABLE);
902 	else
903 		BGE_CLRBIT(sc, BGE_TX_MODE, BGE_TXMODE_FLOWCTL_ENABLE);
904 }
905 
906 /*
907  * Update rx threshold levels to values in a particular slot
908  * of the interrupt-mitigation table bge_rx_threshes.
909  */
910 static void
911 bge_set_thresh(struct ifnet *ifp, int lvl)
912 {
913 	struct bge_softc *sc = ifp->if_softc;
914 	int s;
915 
916 	/* For now, just save the new Rx-intr thresholds and record
917 	 * that a threshold update is pending.  Updating the hardware
918 	 * registers here (even at splhigh()) is observed to
919 	 * occasionaly cause glitches where Rx-interrupts are not
920 	 * honoured for up to 10 seconds. jonathan@NetBSD.org, 2003-04-05
921 	 */
922 	s = splnet();
923 	sc->bge_rx_coal_ticks = bge_rx_threshes[lvl].rx_ticks;
924 	sc->bge_rx_max_coal_bds = bge_rx_threshes[lvl].rx_max_bds;
925 	sc->bge_pending_rxintr_change = 1;
926 	splx(s);
927 
928 	 return;
929 }
930 
931 
932 /*
933  * Update Rx thresholds of all bge devices
934  */
935 static void
936 bge_update_all_threshes(int lvl)
937 {
938 	struct ifnet *ifp;
939 	const char * const namebuf = "bge";
940 	int namelen;
941 
942 	if (lvl < 0)
943 		lvl = 0;
944 	else if( lvl >= NBGE_RX_THRESH)
945 		lvl = NBGE_RX_THRESH - 1;
946 
947 	namelen = strlen(namebuf);
948 	/*
949 	 * Now search all the interfaces for this name/number
950 	 */
951 	IFNET_FOREACH(ifp) {
952 		if (strncmp(ifp->if_xname, namebuf, namelen) != 0)
953 		      continue;
954 		/* We got a match: update if doing auto-threshold-tuning */
955 		if (bge_auto_thresh)
956 			bge_set_thresh(ifp, lvl);
957 	}
958 }
959 
960 /*
961  * Handle events that have triggered interrupts.
962  */
963 static void
964 bge_handle_events(struct bge_softc *sc)
965 {
966 
967 	return;
968 }
969 
970 /*
971  * Memory management for jumbo frames.
972  */
973 
974 static int
975 bge_alloc_jumbo_mem(struct bge_softc *sc)
976 {
977 	char *ptr, *kva;
978 	bus_dma_segment_t	seg;
979 	int		i, rseg, state, error;
980 	struct bge_jpool_entry   *entry;
981 
982 	state = error = 0;
983 
984 	/* Grab a big chunk o' storage. */
985 	if (bus_dmamem_alloc(sc->bge_dmatag, BGE_JMEM, PAGE_SIZE, 0,
986 	     &seg, 1, &rseg, BUS_DMA_NOWAIT)) {
987 		aprint_error_dev(sc->bge_dev, "can't alloc rx buffers\n");
988 		return ENOBUFS;
989 	}
990 
991 	state = 1;
992 	if (bus_dmamem_map(sc->bge_dmatag, &seg, rseg, BGE_JMEM, (void **)&kva,
993 	    BUS_DMA_NOWAIT)) {
994 		aprint_error_dev(sc->bge_dev,
995 		    "can't map DMA buffers (%d bytes)\n", (int)BGE_JMEM);
996 		error = ENOBUFS;
997 		goto out;
998 	}
999 
1000 	state = 2;
1001 	if (bus_dmamap_create(sc->bge_dmatag, BGE_JMEM, 1, BGE_JMEM, 0,
1002 	    BUS_DMA_NOWAIT, &sc->bge_cdata.bge_rx_jumbo_map)) {
1003 		aprint_error_dev(sc->bge_dev, "can't create DMA map\n");
1004 		error = ENOBUFS;
1005 		goto out;
1006 	}
1007 
1008 	state = 3;
1009 	if (bus_dmamap_load(sc->bge_dmatag, sc->bge_cdata.bge_rx_jumbo_map,
1010 	    kva, BGE_JMEM, NULL, BUS_DMA_NOWAIT)) {
1011 		aprint_error_dev(sc->bge_dev, "can't load DMA map\n");
1012 		error = ENOBUFS;
1013 		goto out;
1014 	}
1015 
1016 	state = 4;
1017 	sc->bge_cdata.bge_jumbo_buf = (void *)kva;
1018 	DPRINTFN(1,("bge_jumbo_buf = %p\n", sc->bge_cdata.bge_jumbo_buf));
1019 
1020 	SLIST_INIT(&sc->bge_jfree_listhead);
1021 	SLIST_INIT(&sc->bge_jinuse_listhead);
1022 
1023 	/*
1024 	 * Now divide it up into 9K pieces and save the addresses
1025 	 * in an array.
1026 	 */
1027 	ptr = sc->bge_cdata.bge_jumbo_buf;
1028 	for (i = 0; i < BGE_JSLOTS; i++) {
1029 		sc->bge_cdata.bge_jslots[i] = ptr;
1030 		ptr += BGE_JLEN;
1031 		entry = malloc(sizeof(struct bge_jpool_entry),
1032 		    M_DEVBUF, M_NOWAIT);
1033 		if (entry == NULL) {
1034 			aprint_error_dev(sc->bge_dev,
1035 			    "no memory for jumbo buffer queue!\n");
1036 			error = ENOBUFS;
1037 			goto out;
1038 		}
1039 		entry->slot = i;
1040 		SLIST_INSERT_HEAD(&sc->bge_jfree_listhead,
1041 				 entry, jpool_entries);
1042 	}
1043 out:
1044 	if (error != 0) {
1045 		switch (state) {
1046 		case 4:
1047 			bus_dmamap_unload(sc->bge_dmatag,
1048 			    sc->bge_cdata.bge_rx_jumbo_map);
1049 		case 3:
1050 			bus_dmamap_destroy(sc->bge_dmatag,
1051 			    sc->bge_cdata.bge_rx_jumbo_map);
1052 		case 2:
1053 			bus_dmamem_unmap(sc->bge_dmatag, kva, BGE_JMEM);
1054 		case 1:
1055 			bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
1056 			break;
1057 		default:
1058 			break;
1059 		}
1060 	}
1061 
1062 	return error;
1063 }
1064 
1065 /*
1066  * Allocate a jumbo buffer.
1067  */
1068 static void *
1069 bge_jalloc(struct bge_softc *sc)
1070 {
1071 	struct bge_jpool_entry   *entry;
1072 
1073 	entry = SLIST_FIRST(&sc->bge_jfree_listhead);
1074 
1075 	if (entry == NULL) {
1076 		aprint_error_dev(sc->bge_dev, "no free jumbo buffers\n");
1077 		return (NULL);
1078 	}
1079 
1080 	SLIST_REMOVE_HEAD(&sc->bge_jfree_listhead, jpool_entries);
1081 	SLIST_INSERT_HEAD(&sc->bge_jinuse_listhead, entry, jpool_entries);
1082 	return (sc->bge_cdata.bge_jslots[entry->slot]);
1083 }
1084 
1085 /*
1086  * Release a jumbo buffer.
1087  */
1088 static void
1089 bge_jfree(struct mbuf *m, void *buf, size_t size, void *arg)
1090 {
1091 	struct bge_jpool_entry *entry;
1092 	struct bge_softc *sc;
1093 	int i, s;
1094 
1095 	/* Extract the softc struct pointer. */
1096 	sc = (struct bge_softc *)arg;
1097 
1098 	if (sc == NULL)
1099 		panic("bge_jfree: can't find softc pointer!");
1100 
1101 	/* calculate the slot this buffer belongs to */
1102 
1103 	i = ((char *)buf
1104 	     - (char *)sc->bge_cdata.bge_jumbo_buf) / BGE_JLEN;
1105 
1106 	if ((i < 0) || (i >= BGE_JSLOTS))
1107 		panic("bge_jfree: asked to free buffer that we don't manage!");
1108 
1109 	s = splvm();
1110 	entry = SLIST_FIRST(&sc->bge_jinuse_listhead);
1111 	if (entry == NULL)
1112 		panic("bge_jfree: buffer not in use!");
1113 	entry->slot = i;
1114 	SLIST_REMOVE_HEAD(&sc->bge_jinuse_listhead, jpool_entries);
1115 	SLIST_INSERT_HEAD(&sc->bge_jfree_listhead, entry, jpool_entries);
1116 
1117 	if (__predict_true(m != NULL))
1118   		pool_cache_put(mb_cache, m);
1119 	splx(s);
1120 }
1121 
1122 
1123 /*
1124  * Intialize a standard receive ring descriptor.
1125  */
1126 static int
1127 bge_newbuf_std(struct bge_softc *sc, int i, struct mbuf *m, bus_dmamap_t dmamap)
1128 {
1129 	struct mbuf		*m_new = NULL;
1130 	struct bge_rx_bd	*r;
1131 	int			error;
1132 
1133 	if (dmamap == NULL) {
1134 		error = bus_dmamap_create(sc->bge_dmatag, MCLBYTES, 1,
1135 		    MCLBYTES, 0, BUS_DMA_NOWAIT, &dmamap);
1136 		if (error != 0)
1137 			return error;
1138 	}
1139 
1140 	sc->bge_cdata.bge_rx_std_map[i] = dmamap;
1141 
1142 	if (m == NULL) {
1143 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
1144 		if (m_new == NULL)
1145 			return (ENOBUFS);
1146 
1147 		MCLGET(m_new, M_DONTWAIT);
1148 		if (!(m_new->m_flags & M_EXT)) {
1149 			m_freem(m_new);
1150 			return (ENOBUFS);
1151 		}
1152 		m_new->m_len = m_new->m_pkthdr.len = MCLBYTES;
1153 
1154 	} else {
1155 		m_new = m;
1156 		m_new->m_len = m_new->m_pkthdr.len = MCLBYTES;
1157 		m_new->m_data = m_new->m_ext.ext_buf;
1158 	}
1159 	if (!(sc->bge_flags & BGE_RX_ALIGNBUG))
1160 	    m_adj(m_new, ETHER_ALIGN);
1161 	if (bus_dmamap_load_mbuf(sc->bge_dmatag, dmamap, m_new,
1162 	    BUS_DMA_READ|BUS_DMA_NOWAIT))
1163 		return (ENOBUFS);
1164 	bus_dmamap_sync(sc->bge_dmatag, dmamap, 0, dmamap->dm_mapsize,
1165 	    BUS_DMASYNC_PREREAD);
1166 
1167 	sc->bge_cdata.bge_rx_std_chain[i] = m_new;
1168 	r = &sc->bge_rdata->bge_rx_std_ring[i];
1169 	bge_set_hostaddr(&r->bge_addr,
1170 	    dmamap->dm_segs[0].ds_addr);
1171 	r->bge_flags = BGE_RXBDFLAG_END;
1172 	r->bge_len = m_new->m_len;
1173 	r->bge_idx = i;
1174 
1175 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
1176 	    offsetof(struct bge_ring_data, bge_rx_std_ring) +
1177 		i * sizeof (struct bge_rx_bd),
1178 	    sizeof (struct bge_rx_bd),
1179 	    BUS_DMASYNC_PREWRITE|BUS_DMASYNC_PREREAD);
1180 
1181 	return (0);
1182 }
1183 
1184 /*
1185  * Initialize a jumbo receive ring descriptor. This allocates
1186  * a jumbo buffer from the pool managed internally by the driver.
1187  */
1188 static int
1189 bge_newbuf_jumbo(struct bge_softc *sc, int i, struct mbuf *m)
1190 {
1191 	struct mbuf *m_new = NULL;
1192 	struct bge_rx_bd *r;
1193 	void *buf = NULL;
1194 
1195 	if (m == NULL) {
1196 
1197 		/* Allocate the mbuf. */
1198 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
1199 		if (m_new == NULL)
1200 			return (ENOBUFS);
1201 
1202 		/* Allocate the jumbo buffer */
1203 		buf = bge_jalloc(sc);
1204 		if (buf == NULL) {
1205 			m_freem(m_new);
1206 			aprint_error_dev(sc->bge_dev,
1207 			    "jumbo allocation failed -- packet dropped!\n");
1208 			return (ENOBUFS);
1209 		}
1210 
1211 		/* Attach the buffer to the mbuf. */
1212 		m_new->m_len = m_new->m_pkthdr.len = BGE_JUMBO_FRAMELEN;
1213 		MEXTADD(m_new, buf, BGE_JUMBO_FRAMELEN, M_DEVBUF,
1214 		    bge_jfree, sc);
1215 		m_new->m_flags |= M_EXT_RW;
1216 	} else {
1217 		m_new = m;
1218 		buf = m_new->m_data = m_new->m_ext.ext_buf;
1219 		m_new->m_ext.ext_size = BGE_JUMBO_FRAMELEN;
1220 	}
1221 	if (!(sc->bge_flags & BGE_RX_ALIGNBUG))
1222 	    m_adj(m_new, ETHER_ALIGN);
1223 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_cdata.bge_rx_jumbo_map,
1224 	    mtod(m_new, char *) - (char *)sc->bge_cdata.bge_jumbo_buf, BGE_JLEN,
1225 	    BUS_DMASYNC_PREREAD);
1226 	/* Set up the descriptor. */
1227 	r = &sc->bge_rdata->bge_rx_jumbo_ring[i];
1228 	sc->bge_cdata.bge_rx_jumbo_chain[i] = m_new;
1229 	bge_set_hostaddr(&r->bge_addr, BGE_JUMBO_DMA_ADDR(sc, m_new));
1230 	r->bge_flags = BGE_RXBDFLAG_END|BGE_RXBDFLAG_JUMBO_RING;
1231 	r->bge_len = m_new->m_len;
1232 	r->bge_idx = i;
1233 
1234 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
1235 	    offsetof(struct bge_ring_data, bge_rx_jumbo_ring) +
1236 		i * sizeof (struct bge_rx_bd),
1237 	    sizeof (struct bge_rx_bd),
1238 	    BUS_DMASYNC_PREWRITE|BUS_DMASYNC_PREREAD);
1239 
1240 	return (0);
1241 }
1242 
1243 /*
1244  * The standard receive ring has 512 entries in it. At 2K per mbuf cluster,
1245  * that's 1MB or memory, which is a lot. For now, we fill only the first
1246  * 256 ring entries and hope that our CPU is fast enough to keep up with
1247  * the NIC.
1248  */
1249 static int
1250 bge_init_rx_ring_std(struct bge_softc *sc)
1251 {
1252 	int i;
1253 
1254 	if (sc->bge_flags & BGE_RXRING_VALID)
1255 		return 0;
1256 
1257 	for (i = 0; i < BGE_SSLOTS; i++) {
1258 		if (bge_newbuf_std(sc, i, NULL, 0) == ENOBUFS)
1259 			return (ENOBUFS);
1260 	}
1261 
1262 	sc->bge_std = i - 1;
1263 	bge_writembx(sc, BGE_MBX_RX_STD_PROD_LO, sc->bge_std);
1264 
1265 	sc->bge_flags |= BGE_RXRING_VALID;
1266 
1267 	return (0);
1268 }
1269 
1270 static void
1271 bge_free_rx_ring_std(struct bge_softc *sc)
1272 {
1273 	int i;
1274 
1275 	if (!(sc->bge_flags & BGE_RXRING_VALID))
1276 		return;
1277 
1278 	for (i = 0; i < BGE_STD_RX_RING_CNT; i++) {
1279 		if (sc->bge_cdata.bge_rx_std_chain[i] != NULL) {
1280 			m_freem(sc->bge_cdata.bge_rx_std_chain[i]);
1281 			sc->bge_cdata.bge_rx_std_chain[i] = NULL;
1282 			bus_dmamap_destroy(sc->bge_dmatag,
1283 			    sc->bge_cdata.bge_rx_std_map[i]);
1284 		}
1285 		memset((char *)&sc->bge_rdata->bge_rx_std_ring[i], 0,
1286 		    sizeof(struct bge_rx_bd));
1287 	}
1288 
1289 	sc->bge_flags &= ~BGE_RXRING_VALID;
1290 }
1291 
1292 static int
1293 bge_init_rx_ring_jumbo(struct bge_softc *sc)
1294 {
1295 	int i;
1296 	volatile struct bge_rcb *rcb;
1297 
1298 	if (sc->bge_flags & BGE_JUMBO_RXRING_VALID)
1299 		return 0;
1300 
1301 	for (i = 0; i < BGE_JUMBO_RX_RING_CNT; i++) {
1302 		if (bge_newbuf_jumbo(sc, i, NULL) == ENOBUFS)
1303 			return (ENOBUFS);
1304 	};
1305 
1306 	sc->bge_jumbo = i - 1;
1307 	sc->bge_flags |= BGE_JUMBO_RXRING_VALID;
1308 
1309 	rcb = &sc->bge_rdata->bge_info.bge_jumbo_rx_rcb;
1310 	rcb->bge_maxlen_flags = 0;
1311 	CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_MAXLEN_FLAGS, rcb->bge_maxlen_flags);
1312 
1313 	bge_writembx(sc, BGE_MBX_RX_JUMBO_PROD_LO, sc->bge_jumbo);
1314 
1315 	return (0);
1316 }
1317 
1318 static void
1319 bge_free_rx_ring_jumbo(struct bge_softc *sc)
1320 {
1321 	int i;
1322 
1323 	if (!(sc->bge_flags & BGE_JUMBO_RXRING_VALID))
1324 		return;
1325 
1326 	for (i = 0; i < BGE_JUMBO_RX_RING_CNT; i++) {
1327 		if (sc->bge_cdata.bge_rx_jumbo_chain[i] != NULL) {
1328 			m_freem(sc->bge_cdata.bge_rx_jumbo_chain[i]);
1329 			sc->bge_cdata.bge_rx_jumbo_chain[i] = NULL;
1330 		}
1331 		memset((char *)&sc->bge_rdata->bge_rx_jumbo_ring[i], 0,
1332 		    sizeof(struct bge_rx_bd));
1333 	}
1334 
1335 	sc->bge_flags &= ~BGE_JUMBO_RXRING_VALID;
1336 }
1337 
1338 static void
1339 bge_free_tx_ring(struct bge_softc *sc)
1340 {
1341 	int i, freed;
1342 	struct txdmamap_pool_entry *dma;
1343 
1344 	if (!(sc->bge_flags & BGE_TXRING_VALID))
1345 		return;
1346 
1347 	freed = 0;
1348 
1349 	for (i = 0; i < BGE_TX_RING_CNT; i++) {
1350 		if (sc->bge_cdata.bge_tx_chain[i] != NULL) {
1351 			freed++;
1352 			m_freem(sc->bge_cdata.bge_tx_chain[i]);
1353 			sc->bge_cdata.bge_tx_chain[i] = NULL;
1354 			SLIST_INSERT_HEAD(&sc->txdma_list, sc->txdma[i],
1355 					    link);
1356 			sc->txdma[i] = 0;
1357 		}
1358 		memset((char *)&sc->bge_rdata->bge_tx_ring[i], 0,
1359 		    sizeof(struct bge_tx_bd));
1360 	}
1361 
1362 	while ((dma = SLIST_FIRST(&sc->txdma_list))) {
1363 		SLIST_REMOVE_HEAD(&sc->txdma_list, link);
1364 		bus_dmamap_destroy(sc->bge_dmatag, dma->dmamap);
1365 		free(dma, M_DEVBUF);
1366 	}
1367 
1368 	sc->bge_flags &= ~BGE_TXRING_VALID;
1369 }
1370 
1371 static int
1372 bge_init_tx_ring(struct bge_softc *sc)
1373 {
1374 	int i;
1375 	bus_dmamap_t dmamap;
1376 	struct txdmamap_pool_entry *dma;
1377 
1378 	if (sc->bge_flags & BGE_TXRING_VALID)
1379 		return 0;
1380 
1381 	sc->bge_txcnt = 0;
1382 	sc->bge_tx_saved_considx = 0;
1383 
1384 	/* Initialize transmit producer index for host-memory send ring. */
1385 	sc->bge_tx_prodidx = 0;
1386 	bge_writembx(sc, BGE_MBX_TX_HOST_PROD0_LO, sc->bge_tx_prodidx);
1387 	/* 5700 b2 errata */
1388 	if (BGE_CHIPREV(sc->bge_chipid) == BGE_CHIPREV_5700_BX)
1389 		bge_writembx(sc, BGE_MBX_TX_HOST_PROD0_LO, sc->bge_tx_prodidx);
1390 
1391 	/* NIC-memory send ring not used; initialize to zero. */
1392 	bge_writembx(sc, BGE_MBX_TX_NIC_PROD0_LO, 0);
1393 	/* 5700 b2 errata */
1394 	if (BGE_CHIPREV(sc->bge_chipid) == BGE_CHIPREV_5700_BX)
1395 		bge_writembx(sc, BGE_MBX_TX_NIC_PROD0_LO, 0);
1396 
1397 	SLIST_INIT(&sc->txdma_list);
1398 	for (i = 0; i < BGE_RSLOTS; i++) {
1399 		if (bus_dmamap_create(sc->bge_dmatag, BGE_TXDMA_MAX,
1400 		    BGE_NTXSEG, ETHER_MAX_LEN_JUMBO, 0, BUS_DMA_NOWAIT,
1401 		    &dmamap))
1402 			return (ENOBUFS);
1403 		if (dmamap == NULL)
1404 			panic("dmamap NULL in bge_init_tx_ring");
1405 		dma = malloc(sizeof(*dma), M_DEVBUF, M_NOWAIT);
1406 		if (dma == NULL) {
1407 			aprint_error_dev(sc->bge_dev,
1408 			    "can't alloc txdmamap_pool_entry\n");
1409 			bus_dmamap_destroy(sc->bge_dmatag, dmamap);
1410 			return (ENOMEM);
1411 		}
1412 		dma->dmamap = dmamap;
1413 		SLIST_INSERT_HEAD(&sc->txdma_list, dma, link);
1414 	}
1415 
1416 	sc->bge_flags |= BGE_TXRING_VALID;
1417 
1418 	return (0);
1419 }
1420 
1421 static void
1422 bge_setmulti(struct bge_softc *sc)
1423 {
1424 	struct ethercom		*ac = &sc->ethercom;
1425 	struct ifnet		*ifp = &ac->ec_if;
1426 	struct ether_multi	*enm;
1427 	struct ether_multistep  step;
1428 	u_int32_t		hashes[4] = { 0, 0, 0, 0 };
1429 	u_int32_t		h;
1430 	int			i;
1431 
1432 	if (ifp->if_flags & IFF_PROMISC)
1433 		goto allmulti;
1434 
1435 	/* Now program new ones. */
1436 	ETHER_FIRST_MULTI(step, ac, enm);
1437 	while (enm != NULL) {
1438 		if (memcmp(enm->enm_addrlo, enm->enm_addrhi, ETHER_ADDR_LEN)) {
1439 			/*
1440 			 * We must listen to a range of multicast addresses.
1441 			 * For now, just accept all multicasts, rather than
1442 			 * trying to set only those filter bits needed to match
1443 			 * the range.  (At this time, the only use of address
1444 			 * ranges is for IP multicast routing, for which the
1445 			 * range is big enough to require all bits set.)
1446 			 */
1447 			goto allmulti;
1448 		}
1449 
1450 		h = ether_crc32_le(enm->enm_addrlo, ETHER_ADDR_LEN);
1451 
1452 		/* Just want the 7 least-significant bits. */
1453 		h &= 0x7f;
1454 
1455 		hashes[(h & 0x60) >> 5] |= 1 << (h & 0x1F);
1456 		ETHER_NEXT_MULTI(step, enm);
1457 	}
1458 
1459 	ifp->if_flags &= ~IFF_ALLMULTI;
1460 	goto setit;
1461 
1462  allmulti:
1463 	ifp->if_flags |= IFF_ALLMULTI;
1464 	hashes[0] = hashes[1] = hashes[2] = hashes[3] = 0xffffffff;
1465 
1466  setit:
1467 	for (i = 0; i < 4; i++)
1468 		CSR_WRITE_4(sc, BGE_MAR0 + (i * 4), hashes[i]);
1469 }
1470 
1471 const int bge_swapbits[] = {
1472 	0,
1473 	BGE_MODECTL_BYTESWAP_DATA,
1474 	BGE_MODECTL_WORDSWAP_DATA,
1475 	BGE_MODECTL_BYTESWAP_NONFRAME,
1476 	BGE_MODECTL_WORDSWAP_NONFRAME,
1477 
1478 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA,
1479 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME,
1480 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_NONFRAME,
1481 
1482 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME,
1483 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_WORDSWAP_NONFRAME,
1484 
1485 	BGE_MODECTL_BYTESWAP_NONFRAME|BGE_MODECTL_WORDSWAP_NONFRAME,
1486 
1487 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
1488 	    BGE_MODECTL_BYTESWAP_NONFRAME,
1489 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
1490 	    BGE_MODECTL_WORDSWAP_NONFRAME,
1491 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME|
1492 	    BGE_MODECTL_WORDSWAP_NONFRAME,
1493 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME|
1494 	    BGE_MODECTL_WORDSWAP_NONFRAME,
1495 
1496 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
1497 	    BGE_MODECTL_BYTESWAP_NONFRAME|BGE_MODECTL_WORDSWAP_NONFRAME,
1498 };
1499 
1500 int bge_swapindex = 0;
1501 
1502 /*
1503  * Do endian, PCI and DMA initialization. Also check the on-board ROM
1504  * self-test results.
1505  */
1506 static int
1507 bge_chipinit(struct bge_softc *sc)
1508 {
1509 	u_int32_t		cachesize;
1510 	int			i;
1511 	u_int32_t		dma_rw_ctl;
1512 
1513 
1514 	/* Set endianness before we access any non-PCI registers. */
1515 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MISC_CTL,
1516 	    BGE_INIT);
1517 
1518 	/* Set power state to D0. */
1519 	bge_setpowerstate(sc, 0);
1520 
1521 	/*
1522 	 * Check the 'ROM failed' bit on the RX CPU to see if
1523 	 * self-tests passed.
1524 	 */
1525 	if (CSR_READ_4(sc, BGE_RXCPU_MODE) & BGE_RXCPUMODE_ROMFAIL) {
1526 		aprint_error_dev(sc->bge_dev,
1527 		    "RX CPU self-diagnostics failed!\n");
1528 		return (ENODEV);
1529 	}
1530 
1531 	/* Clear the MAC control register */
1532 	CSR_WRITE_4(sc, BGE_MAC_MODE, 0);
1533 
1534 	/*
1535 	 * Clear the MAC statistics block in the NIC's
1536 	 * internal memory.
1537 	 */
1538 	for (i = BGE_STATS_BLOCK;
1539 	    i < BGE_STATS_BLOCK_END + 1; i += sizeof(u_int32_t))
1540 		BGE_MEMWIN_WRITE(sc->sc_pc, sc->sc_pcitag, i, 0);
1541 
1542 	for (i = BGE_STATUS_BLOCK;
1543 	    i < BGE_STATUS_BLOCK_END + 1; i += sizeof(u_int32_t))
1544 		BGE_MEMWIN_WRITE(sc->sc_pc, sc->sc_pcitag, i, 0);
1545 
1546 	/* Set up the PCI DMA control register. */
1547 	if (sc->bge_flags & BGE_PCIE) {
1548 	  u_int32_t device_ctl;
1549 
1550 		/* From FreeBSD */
1551 		DPRINTFN(4, ("(%s: PCI-Express DMA setting)\n",
1552 		    device_xname(sc->bge_dev)));
1553 		dma_rw_ctl = (BGE_PCI_READ_CMD | BGE_PCI_WRITE_CMD |
1554 		    (0xf << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
1555 		    (0x2 << BGE_PCIDMARWCTL_WR_WAT_SHIFT));
1556 
1557 		/* jonathan: alternative from Linux driver */
1558 #define DMA_CTRL_WRITE_PCIE_H20MARK_128         0x00180000
1559 #define DMA_CTRL_WRITE_PCIE_H20MARK_256         0x00380000
1560 
1561 		dma_rw_ctl =   0x76000000; /* XXX XXX XXX */;
1562 		device_ctl = pci_conf_read(sc->sc_pc, sc->sc_pcitag,
1563 					   BGE_PCI_CONF_DEV_CTRL);
1564 		aprint_debug_dev(sc->bge_dev, "pcie mode=0x%x\n", device_ctl);
1565 
1566 		if ((device_ctl & 0x00e0) && 0) {
1567 			/*
1568 			 * XXX jonathan@NetBSD.org:
1569 			 * This clause is exactly what the Broadcom-supplied
1570 			 * Linux does; but given overall register programming
1571 			 * by if_bge(4), this larger DMA-write watermark
1572 			 * value causes bcm5721 chips to totally wedge.
1573 			 */
1574 			dma_rw_ctl |= BGE_PCIDMA_RWCTL_PCIE_WRITE_WATRMARK_256;
1575 		} else {
1576 			dma_rw_ctl |= BGE_PCIDMA_RWCTL_PCIE_WRITE_WATRMARK_128;
1577 		}
1578 	} else if (sc->bge_flags & BGE_PCIX){
1579 	  	DPRINTFN(4, ("(:%s: PCI-X DMA setting)\n",
1580 		    device_xname(sc->bge_dev)));
1581 		/* PCI-X bus */
1582 		dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD |
1583 		    (0x3 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
1584 		    (0x3 << BGE_PCIDMARWCTL_WR_WAT_SHIFT) |
1585 		    (0x0F);
1586 		/*
1587 		 * 5703 and 5704 need ONEDMA_AT_ONCE as a workaround
1588 		 * for hardware bugs, which means we should also clear
1589 		 * the low-order MINDMA bits.  In addition, the 5704
1590 		 * uses a different encoding of read/write watermarks.
1591 		 */
1592 		if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5704) {
1593 			dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD |
1594 			  /* should be 0x1f0000 */
1595 			  (0x7 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
1596 			  (0x3 << BGE_PCIDMARWCTL_WR_WAT_SHIFT);
1597 			dma_rw_ctl |= BGE_PCIDMARWCTL_ONEDMA_ATONCE;
1598 		}
1599 		else if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5703) {
1600 			dma_rw_ctl &=  0xfffffff0;
1601 			dma_rw_ctl |= BGE_PCIDMARWCTL_ONEDMA_ATONCE;
1602 		}
1603 		else if (BGE_IS_5714_FAMILY(sc)) {
1604 			dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD;
1605 			dma_rw_ctl &= ~BGE_PCIDMARWCTL_ONEDMA_ATONCE; /* XXX */
1606 			/* XXX magic values, Broadcom-supplied Linux driver */
1607 			if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5780)
1608 				dma_rw_ctl |= (1 << 20) | (1 << 18) |
1609 				  BGE_PCIDMARWCTL_ONEDMA_ATONCE;
1610 			else
1611 				dma_rw_ctl |= (1<<20) | (1<<18) | (1 << 15);
1612 		}
1613 	} else {
1614 		/* Conventional PCI bus */
1615 	  	DPRINTFN(4, ("(%s: PCI 2.2 DMA setting)\n",
1616 		    device_xname(sc->bge_dev)));
1617 		dma_rw_ctl = (BGE_PCI_READ_CMD | BGE_PCI_WRITE_CMD |
1618 		   (0x7 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
1619 		   (0x7 << BGE_PCIDMARWCTL_WR_WAT_SHIFT));
1620 		if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5703 ||
1621 		    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5704)
1622 			dma_rw_ctl |= 0x0F;
1623 	}
1624 
1625 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL, dma_rw_ctl);
1626 
1627 	/*
1628 	 * Set up general mode register.
1629 	 */
1630 	CSR_WRITE_4(sc, BGE_MODE_CTL, BGE_DMA_SWAP_OPTIONS|
1631 		    BGE_MODECTL_MAC_ATTN_INTR|BGE_MODECTL_HOST_SEND_BDS|
1632 		    BGE_MODECTL_TX_NO_PHDR_CSUM|BGE_MODECTL_RX_NO_PHDR_CSUM);
1633 
1634 	/* Get cache line size. */
1635 	cachesize = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CACHESZ);
1636 
1637 	/*
1638 	 * Avoid violating PCI spec on certain chip revs.
1639 	 */
1640 	if (pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CMD) &
1641 	    PCIM_CMD_MWIEN) {
1642 		switch(cachesize) {
1643 		case 1:
1644 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1645 				   BGE_PCI_WRITE_BNDRY_16BYTES);
1646 			break;
1647 		case 2:
1648 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1649 				   BGE_PCI_WRITE_BNDRY_32BYTES);
1650 			break;
1651 		case 4:
1652 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1653 				   BGE_PCI_WRITE_BNDRY_64BYTES);
1654 			break;
1655 		case 8:
1656 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1657 				   BGE_PCI_WRITE_BNDRY_128BYTES);
1658 			break;
1659 		case 16:
1660 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1661 				   BGE_PCI_WRITE_BNDRY_256BYTES);
1662 			break;
1663 		case 32:
1664 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1665 				   BGE_PCI_WRITE_BNDRY_512BYTES);
1666 			break;
1667 		case 64:
1668 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1669 				   BGE_PCI_WRITE_BNDRY_1024BYTES);
1670 			break;
1671 		default:
1672 		/* Disable PCI memory write and invalidate. */
1673 #if 0
1674 			if (bootverbose)
1675 				aprint_error_dev(sc->bge_dev,
1676 				    "cache line size %d not supported "
1677 				    "disabling PCI MWI\n",
1678 #endif
1679 			PCI_CLRBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CMD,
1680 			    PCIM_CMD_MWIEN);
1681 			break;
1682 		}
1683 	}
1684 
1685 	/*
1686 	 * Disable memory write invalidate.  Apparently it is not supported
1687 	 * properly by these devices.
1688 	 */
1689 	PCI_CLRBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CMD, PCIM_CMD_MWIEN);
1690 
1691 
1692 #ifdef __brokenalpha__
1693 	/*
1694 	 * Must insure that we do not cross an 8K (bytes) boundary
1695 	 * for DMA reads.  Our highest limit is 1K bytes.  This is a
1696 	 * restriction on some ALPHA platforms with early revision
1697 	 * 21174 PCI chipsets, such as the AlphaPC 164lx
1698 	 */
1699 	PCI_SETBIT(sc, BGE_PCI_DMA_RW_CTL, BGE_PCI_READ_BNDRY_1024, 4);
1700 #endif
1701 
1702 	/* Set the timer prescaler (always 66MHz) */
1703 	CSR_WRITE_4(sc, BGE_MISC_CFG, 65 << 1/*BGE_32BITTIME_66MHZ*/);
1704 
1705 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5906) {
1706 		DELAY(40);	/* XXX */
1707 
1708 		/* Put PHY into ready state */
1709 		BGE_CLRBIT(sc, BGE_MISC_CFG, BGE_MISCCFG_EPHY_IDDQ);
1710 		CSR_READ_4(sc, BGE_MISC_CFG); /* Flush */
1711 		DELAY(40);
1712 	}
1713 
1714 	return (0);
1715 }
1716 
1717 static int
1718 bge_blockinit(struct bge_softc *sc)
1719 {
1720 	volatile struct bge_rcb		*rcb;
1721 	bus_size_t		rcb_addr;
1722 	int			i;
1723 	struct ifnet		*ifp = &sc->ethercom.ec_if;
1724 	bge_hostaddr		taddr;
1725 
1726 	/*
1727 	 * Initialize the memory window pointer register so that
1728 	 * we can access the first 32K of internal NIC RAM. This will
1729 	 * allow us to set up the TX send ring RCBs and the RX return
1730 	 * ring RCBs, plus other things which live in NIC memory.
1731 	 */
1732 
1733 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MEMWIN_BASEADDR, 0);
1734 
1735 	/* Configure mbuf memory pool */
1736 	if (!(BGE_IS_5705_OR_BEYOND(sc))) {
1737 		if (sc->bge_extram) {
1738 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_BASEADDR,
1739 			    BGE_EXT_SSRAM);
1740 			if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5704)
1741 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x10000);
1742 			else
1743 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x18000);
1744 		} else {
1745 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_BASEADDR,
1746 			    BGE_BUFFPOOL_1);
1747 			if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5704)
1748 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x10000);
1749 			else
1750 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x18000);
1751 		}
1752 
1753 		/* Configure DMA resource pool */
1754 		CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_BASEADDR,
1755 		    BGE_DMA_DESCRIPTORS);
1756 		CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_LEN, 0x2000);
1757 	}
1758 
1759 	/* Configure mbuf pool watermarks */
1760 #ifdef ORIG_WPAUL_VALUES
1761 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 24);
1762 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 24);
1763 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 48);
1764 #else
1765 
1766 	/* new broadcom docs strongly recommend these: */
1767 	if (!BGE_IS_5705_OR_BEYOND(sc)) {
1768 		if (ifp->if_mtu > ETHER_MAX_LEN) {
1769 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 0x50);
1770 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 0x20);
1771 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 0x60);
1772 		} else {
1773 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 304);
1774 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 152);
1775 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 380);
1776 		}
1777 	} else if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5906) {
1778 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 0x0);
1779 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 0x04);
1780 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 0x10);
1781 	} else {
1782 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 0x0);
1783 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 0x10);
1784 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 0x60);
1785 	}
1786 #endif
1787 
1788 	/* Configure DMA resource watermarks */
1789 	CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_LOWAT, 5);
1790 	CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_HIWAT, 10);
1791 
1792 	/* Enable buffer manager */
1793 	if (!BGE_IS_5705_OR_BEYOND(sc)) {
1794 		CSR_WRITE_4(sc, BGE_BMAN_MODE,
1795 		    BGE_BMANMODE_ENABLE|BGE_BMANMODE_LOMBUF_ATTN);
1796 
1797 		/* Poll for buffer manager start indication */
1798 		for (i = 0; i < BGE_TIMEOUT; i++) {
1799 			if (CSR_READ_4(sc, BGE_BMAN_MODE) & BGE_BMANMODE_ENABLE)
1800 				break;
1801 			DELAY(10);
1802 		}
1803 
1804 		if (i == BGE_TIMEOUT) {
1805 			aprint_error_dev(sc->bge_dev,
1806 			    "buffer manager failed to start\n");
1807 			return (ENXIO);
1808 		}
1809 	}
1810 
1811 	/* Enable flow-through queues */
1812 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0xFFFFFFFF);
1813 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0);
1814 
1815 	/* Wait until queue initialization is complete */
1816 	for (i = 0; i < BGE_TIMEOUT; i++) {
1817 		if (CSR_READ_4(sc, BGE_FTQ_RESET) == 0)
1818 			break;
1819 		DELAY(10);
1820 	}
1821 
1822 	if (i == BGE_TIMEOUT) {
1823 		aprint_error_dev(sc->bge_dev,
1824 		    "flow-through queue init failed\n");
1825 		return (ENXIO);
1826 	}
1827 
1828 	/* Initialize the standard RX ring control block */
1829 	rcb = &sc->bge_rdata->bge_info.bge_std_rx_rcb;
1830 	bge_set_hostaddr(&rcb->bge_hostaddr,
1831 	    BGE_RING_DMA_ADDR(sc, bge_rx_std_ring));
1832 	if (BGE_IS_5705_OR_BEYOND(sc))
1833 		rcb->bge_maxlen_flags = BGE_RCB_MAXLEN_FLAGS(512, 0);
1834 	else
1835 		rcb->bge_maxlen_flags =
1836 		    BGE_RCB_MAXLEN_FLAGS(BGE_MAX_FRAMELEN, 0);
1837 	if (sc->bge_extram)
1838 		rcb->bge_nicaddr = BGE_EXT_STD_RX_RINGS;
1839 	else
1840 		rcb->bge_nicaddr = BGE_STD_RX_RINGS;
1841 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_HADDR_HI, rcb->bge_hostaddr.bge_addr_hi);
1842 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_HADDR_LO, rcb->bge_hostaddr.bge_addr_lo);
1843 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_MAXLEN_FLAGS, rcb->bge_maxlen_flags);
1844 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_NICADDR, rcb->bge_nicaddr);
1845 
1846 	if (BGE_IS_5705_OR_BEYOND(sc))
1847 		sc->bge_return_ring_cnt = BGE_RETURN_RING_CNT_5705;
1848 	else
1849 		sc->bge_return_ring_cnt = BGE_RETURN_RING_CNT;
1850 
1851 	/*
1852 	 * Initialize the jumbo RX ring control block
1853 	 * We set the 'ring disabled' bit in the flags
1854 	 * field until we're actually ready to start
1855 	 * using this ring (i.e. once we set the MTU
1856 	 * high enough to require it).
1857 	 */
1858 	if (!BGE_IS_5705_OR_BEYOND(sc)) {
1859 		rcb = &sc->bge_rdata->bge_info.bge_jumbo_rx_rcb;
1860 		bge_set_hostaddr(&rcb->bge_hostaddr,
1861 		    BGE_RING_DMA_ADDR(sc, bge_rx_jumbo_ring));
1862 		rcb->bge_maxlen_flags =
1863 		    BGE_RCB_MAXLEN_FLAGS(BGE_MAX_FRAMELEN,
1864 			BGE_RCB_FLAG_RING_DISABLED);
1865 		if (sc->bge_extram)
1866 			rcb->bge_nicaddr = BGE_EXT_JUMBO_RX_RINGS;
1867 		else
1868 			rcb->bge_nicaddr = BGE_JUMBO_RX_RINGS;
1869 
1870 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_HADDR_HI,
1871 		    rcb->bge_hostaddr.bge_addr_hi);
1872 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_HADDR_LO,
1873 		    rcb->bge_hostaddr.bge_addr_lo);
1874 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_MAXLEN_FLAGS,
1875 		    rcb->bge_maxlen_flags);
1876 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_NICADDR, rcb->bge_nicaddr);
1877 
1878 		/* Set up dummy disabled mini ring RCB */
1879 		rcb = &sc->bge_rdata->bge_info.bge_mini_rx_rcb;
1880 		rcb->bge_maxlen_flags = BGE_RCB_MAXLEN_FLAGS(0,
1881 		    BGE_RCB_FLAG_RING_DISABLED);
1882 		CSR_WRITE_4(sc, BGE_RX_MINI_RCB_MAXLEN_FLAGS,
1883 		    rcb->bge_maxlen_flags);
1884 
1885 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
1886 		    offsetof(struct bge_ring_data, bge_info),
1887 		    sizeof (struct bge_gib),
1888 		    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1889 	}
1890 
1891 	/*
1892 	 * Set the BD ring replenish thresholds. The recommended
1893 	 * values are 1/8th the number of descriptors allocated to
1894 	 * each ring.
1895 	 */
1896 	i = BGE_STD_RX_RING_CNT / 8;
1897 
1898 	/*
1899 	 * Use a value of 8 for the following chips to workaround HW errata.
1900 	 * Some of these chips have been added based on empirical
1901 	 * evidence (they don't work unless this is done).
1902 	 */
1903 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5750 ||
1904 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5752 ||
1905 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5755 ||
1906 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5787 ||
1907 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5906)
1908 		i = 8;
1909 
1910 	CSR_WRITE_4(sc, BGE_RBDI_STD_REPL_THRESH, i);
1911 	CSR_WRITE_4(sc, BGE_RBDI_JUMBO_REPL_THRESH, BGE_JUMBO_RX_RING_CNT/8);
1912 
1913 	/*
1914 	 * Disable all unused send rings by setting the 'ring disabled'
1915 	 * bit in the flags field of all the TX send ring control blocks.
1916 	 * These are located in NIC memory.
1917 	 */
1918 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
1919 	for (i = 0; i < BGE_TX_RINGS_EXTSSRAM_MAX; i++) {
1920 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1921 		    BGE_RCB_MAXLEN_FLAGS(0, BGE_RCB_FLAG_RING_DISABLED));
1922 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
1923 		rcb_addr += sizeof(struct bge_rcb);
1924 	}
1925 
1926 	/* Configure TX RCB 0 (we use only the first ring) */
1927 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
1928 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_tx_ring));
1929 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
1930 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
1931 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr,
1932 		    BGE_NIC_TXRING_ADDR(0, BGE_TX_RING_CNT));
1933 	if (!(BGE_IS_5705_OR_BEYOND(sc)))
1934 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1935 		    BGE_RCB_MAXLEN_FLAGS(BGE_TX_RING_CNT, 0));
1936 
1937 	/* Disable all unused RX return rings */
1938 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
1939 	for (i = 0; i < BGE_RX_RINGS_MAX; i++) {
1940 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, 0);
1941 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, 0);
1942 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1943 			    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt,
1944                                      BGE_RCB_FLAG_RING_DISABLED));
1945 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
1946 		bge_writembx(sc, BGE_MBX_RX_CONS0_LO +
1947 		    (i * (sizeof(u_int64_t))), 0);
1948 		rcb_addr += sizeof(struct bge_rcb);
1949 	}
1950 
1951 	/* Initialize RX ring indexes */
1952 	bge_writembx(sc, BGE_MBX_RX_STD_PROD_LO, 0);
1953 	bge_writembx(sc, BGE_MBX_RX_JUMBO_PROD_LO, 0);
1954 	bge_writembx(sc, BGE_MBX_RX_MINI_PROD_LO, 0);
1955 
1956 	/*
1957 	 * Set up RX return ring 0
1958 	 * Note that the NIC address for RX return rings is 0x00000000.
1959 	 * The return rings live entirely within the host, so the
1960 	 * nicaddr field in the RCB isn't used.
1961 	 */
1962 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
1963 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_rx_return_ring));
1964 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
1965 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
1966 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0x00000000);
1967 	RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1968 	    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt, 0));
1969 
1970 	/* Set random backoff seed for TX */
1971 	CSR_WRITE_4(sc, BGE_TX_RANDOM_BACKOFF,
1972 	    CLLADDR(ifp->if_sadl)[0] + CLLADDR(ifp->if_sadl)[1] +
1973 	    CLLADDR(ifp->if_sadl)[2] + CLLADDR(ifp->if_sadl)[3] +
1974 	    CLLADDR(ifp->if_sadl)[4] + CLLADDR(ifp->if_sadl)[5] +
1975 	    BGE_TX_BACKOFF_SEED_MASK);
1976 
1977 	/* Set inter-packet gap */
1978 	CSR_WRITE_4(sc, BGE_TX_LENGTHS, 0x2620);
1979 
1980 	/*
1981 	 * Specify which ring to use for packets that don't match
1982 	 * any RX rules.
1983 	 */
1984 	CSR_WRITE_4(sc, BGE_RX_RULES_CFG, 0x08);
1985 
1986 	/*
1987 	 * Configure number of RX lists. One interrupt distribution
1988 	 * list, sixteen active lists, one bad frames class.
1989 	 */
1990 	CSR_WRITE_4(sc, BGE_RXLP_CFG, 0x181);
1991 
1992 	/* Inialize RX list placement stats mask. */
1993 	CSR_WRITE_4(sc, BGE_RXLP_STATS_ENABLE_MASK, 0x007FFFFF);
1994 	CSR_WRITE_4(sc, BGE_RXLP_STATS_CTL, 0x1);
1995 
1996 	/* Disable host coalescing until we get it set up */
1997 	CSR_WRITE_4(sc, BGE_HCC_MODE, 0x00000000);
1998 
1999 	/* Poll to make sure it's shut down. */
2000 	for (i = 0; i < BGE_TIMEOUT; i++) {
2001 		if (!(CSR_READ_4(sc, BGE_HCC_MODE) & BGE_HCCMODE_ENABLE))
2002 			break;
2003 		DELAY(10);
2004 	}
2005 
2006 	if (i == BGE_TIMEOUT) {
2007 		aprint_error_dev(sc->bge_dev,
2008 		    "host coalescing engine failed to idle\n");
2009 		return (ENXIO);
2010 	}
2011 
2012 	/* Set up host coalescing defaults */
2013 	CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, sc->bge_rx_coal_ticks);
2014 	CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS, sc->bge_tx_coal_ticks);
2015 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, sc->bge_rx_max_coal_bds);
2016 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS, sc->bge_tx_max_coal_bds);
2017 	if (!(BGE_IS_5705_OR_BEYOND(sc))) {
2018 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS_INT, 0);
2019 		CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS_INT, 0);
2020 	}
2021 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS_INT, 0);
2022 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS_INT, 0);
2023 
2024 	/* Set up address of statistics block */
2025 	if (!(BGE_IS_5705_OR_BEYOND(sc))) {
2026 		bge_set_hostaddr(&taddr,
2027 		    BGE_RING_DMA_ADDR(sc, bge_info.bge_stats));
2028 		CSR_WRITE_4(sc, BGE_HCC_STATS_TICKS, sc->bge_stat_ticks);
2029 		CSR_WRITE_4(sc, BGE_HCC_STATS_BASEADDR, BGE_STATS_BLOCK);
2030 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_HI, taddr.bge_addr_hi);
2031 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_LO, taddr.bge_addr_lo);
2032 	}
2033 
2034 	/* Set up address of status block */
2035 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_status_block));
2036 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_BASEADDR, BGE_STATUS_BLOCK);
2037 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_HI, taddr.bge_addr_hi);
2038 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_LO, taddr.bge_addr_lo);
2039 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx = 0;
2040 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx = 0;
2041 
2042 	/* Turn on host coalescing state machine */
2043 	CSR_WRITE_4(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
2044 
2045 	/* Turn on RX BD completion state machine and enable attentions */
2046 	CSR_WRITE_4(sc, BGE_RBDC_MODE,
2047 	    BGE_RBDCMODE_ENABLE|BGE_RBDCMODE_ATTN);
2048 
2049 	/* Turn on RX list placement state machine */
2050 	CSR_WRITE_4(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
2051 
2052 	/* Turn on RX list selector state machine. */
2053 	if (!(BGE_IS_5705_OR_BEYOND(sc)))
2054 		CSR_WRITE_4(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
2055 
2056 	/* Turn on DMA, clear stats */
2057 	CSR_WRITE_4(sc, BGE_MAC_MODE, BGE_MACMODE_TXDMA_ENB|
2058 	    BGE_MACMODE_RXDMA_ENB|BGE_MACMODE_RX_STATS_CLEAR|
2059 	    BGE_MACMODE_TX_STATS_CLEAR|BGE_MACMODE_RX_STATS_ENB|
2060 	    BGE_MACMODE_TX_STATS_ENB|BGE_MACMODE_FRMHDR_DMA_ENB|
2061 	    ((sc->bge_flags & BGE_PHY_FIBER_TBI) ? BGE_PORTMODE_TBI
2062 		: BGE_PORTMODE_MII));
2063 
2064 	/* Set misc. local control, enable interrupts on attentions */
2065 	sc->bge_local_ctrl_reg = BGE_MLC_INTR_ONATTN | BGE_MLC_AUTO_EEPROM;
2066 
2067 #ifdef notdef
2068 	/* Assert GPIO pins for PHY reset */
2069 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUT0|
2070 	    BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUT2);
2071 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUTEN0|
2072 	    BGE_MLC_MISCIO_OUTEN1|BGE_MLC_MISCIO_OUTEN2);
2073 #endif
2074 
2075 #if defined(not_quite_yet)
2076 	/* Linux driver enables enable gpio pin #1 on 5700s */
2077 	if (sc->bge_chipid == BGE_CHIPID_BCM5700) {
2078 		sc->bge_local_ctrl_reg |=
2079 		  (BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUTEN1);
2080 	}
2081 #endif
2082 	CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
2083 
2084 	/* Turn on DMA completion state machine */
2085 	if (!(BGE_IS_5705_OR_BEYOND(sc)))
2086 		CSR_WRITE_4(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
2087 
2088 	/* Turn on write DMA state machine */
2089 	{
2090 		uint32_t bge_wdma_mode =
2091 			BGE_WDMAMODE_ENABLE|BGE_WDMAMODE_ALL_ATTNS;
2092 
2093 		/* Enable host coalescing bug fix; see Linux tg3.c */
2094 		if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5755 ||
2095 		    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5787)
2096 			bge_wdma_mode |= (1 << 29);
2097 
2098 		CSR_WRITE_4(sc, BGE_WDMA_MODE, bge_wdma_mode);
2099         }
2100 
2101 	/* Turn on read DMA state machine */
2102 	{
2103 		uint32_t dma_read_modebits;
2104 
2105 		dma_read_modebits =
2106 		  BGE_RDMAMODE_ENABLE | BGE_RDMAMODE_ALL_ATTNS;
2107 
2108 		if ((sc->bge_flags & BGE_PCIE) && 0) {
2109 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_LONG_BURST;
2110 		} else if (BGE_IS_5705_OR_BEYOND(sc)) {
2111 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_SIZE_128;
2112 		}
2113 
2114 		/* XXX broadcom-supplied linux driver; undocumented */
2115 		if (BGE_IS_5750_OR_BEYOND(sc)) {
2116  			/*
2117 			 * XXX: magic values.
2118 			 * From Broadcom-supplied Linux driver;  apparently
2119 			 * required to workaround a DMA bug affecting TSO
2120 			 * on bcm575x/bcm5721?
2121 			 */
2122 			dma_read_modebits |= (1 << 27);
2123 		}
2124 		CSR_WRITE_4(sc, BGE_RDMA_MODE, dma_read_modebits);
2125 	}
2126 
2127 	/* Turn on RX data completion state machine */
2128 	CSR_WRITE_4(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
2129 
2130 	/* Turn on RX BD initiator state machine */
2131 	CSR_WRITE_4(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
2132 
2133 	/* Turn on RX data and RX BD initiator state machine */
2134 	CSR_WRITE_4(sc, BGE_RDBDI_MODE, BGE_RDBDIMODE_ENABLE);
2135 
2136 	/* Turn on Mbuf cluster free state machine */
2137 	if (!(BGE_IS_5705_OR_BEYOND(sc)))
2138 		CSR_WRITE_4(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
2139 
2140 	/* Turn on send BD completion state machine */
2141 	CSR_WRITE_4(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
2142 
2143 	/* Turn on send data completion state machine */
2144 	CSR_WRITE_4(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
2145 
2146 	/* Turn on send data initiator state machine */
2147 	if (BGE_IS_5750_OR_BEYOND(sc)) {
2148 		/* XXX: magic value from Linux driver */
2149 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE | 0x08);
2150 	} else {
2151 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
2152 	}
2153 
2154 	/* Turn on send BD initiator state machine */
2155 	CSR_WRITE_4(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
2156 
2157 	/* Turn on send BD selector state machine */
2158 	CSR_WRITE_4(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
2159 
2160 	CSR_WRITE_4(sc, BGE_SDI_STATS_ENABLE_MASK, 0x007FFFFF);
2161 	CSR_WRITE_4(sc, BGE_SDI_STATS_CTL,
2162 	    BGE_SDISTATSCTL_ENABLE|BGE_SDISTATSCTL_FASTER);
2163 
2164 	/* ack/clear link change events */
2165 	CSR_WRITE_4(sc, BGE_MAC_STS, BGE_MACSTAT_SYNC_CHANGED|
2166 	    BGE_MACSTAT_CFG_CHANGED);
2167 	CSR_WRITE_4(sc, BGE_MI_STS, 0);
2168 
2169 	/* Enable PHY auto polling (for MII/GMII only) */
2170 	if (sc->bge_flags & BGE_PHY_FIBER_TBI) {
2171 		CSR_WRITE_4(sc, BGE_MI_STS, BGE_MISTS_LINK);
2172  	} else {
2173 		BGE_SETBIT(sc, BGE_MI_MODE, BGE_MIMODE_AUTOPOLL|10<<16);
2174 		if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5700)
2175 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
2176 			    BGE_EVTENB_MI_INTERRUPT);
2177 	}
2178 
2179 	/* Enable link state change attentions. */
2180 	BGE_SETBIT(sc, BGE_MAC_EVT_ENB, BGE_EVTENB_LINK_CHANGED);
2181 
2182 	return (0);
2183 }
2184 
2185 static const struct bge_revision *
2186 bge_lookup_rev(uint32_t chipid)
2187 {
2188 	const struct bge_revision *br;
2189 
2190 	for (br = bge_revisions; br->br_name != NULL; br++) {
2191 		if (br->br_chipid == chipid)
2192 			return (br);
2193 	}
2194 
2195 	for (br = bge_majorrevs; br->br_name != NULL; br++) {
2196 		if (br->br_chipid == BGE_ASICREV(chipid))
2197 			return (br);
2198 	}
2199 
2200 	return (NULL);
2201 }
2202 
2203 static const struct bge_product *
2204 bge_lookup(const struct pci_attach_args *pa)
2205 {
2206 	const struct bge_product *bp;
2207 
2208 	for (bp = bge_products; bp->bp_name != NULL; bp++) {
2209 		if (PCI_VENDOR(pa->pa_id) == bp->bp_vendor &&
2210 		    PCI_PRODUCT(pa->pa_id) == bp->bp_product)
2211 			return (bp);
2212 	}
2213 
2214 	return (NULL);
2215 }
2216 
2217 static int
2218 bge_setpowerstate(struct bge_softc *sc, int powerlevel)
2219 {
2220 #ifdef NOTYET
2221 	u_int32_t pm_ctl = 0;
2222 
2223 	/* XXX FIXME: make sure indirect accesses enabled? */
2224 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_MISC_CTL, 4);
2225 	pm_ctl |= BGE_PCIMISCCTL_INDIRECT_ACCESS;
2226 	pci_write_config(sc->bge_dev, BGE_PCI_MISC_CTL, pm_ctl, 4);
2227 
2228 	/* clear the PME_assert bit and power state bits, enable PME */
2229 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_PWRMGMT_CMD, 2);
2230 	pm_ctl &= ~PCIM_PSTAT_DMASK;
2231 	pm_ctl |= (1 << 8);
2232 
2233 	if (powerlevel == 0) {
2234 		pm_ctl |= PCIM_PSTAT_D0;
2235 		pci_write_config(sc->bge_dev, BGE_PCI_PWRMGMT_CMD,
2236 		    pm_ctl, 2);
2237 		DELAY(10000);
2238 		CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
2239 		DELAY(10000);
2240 
2241 #ifdef NOTYET
2242 		/* XXX FIXME: write 0x02 to phy aux_Ctrl reg */
2243 		bge_miibus_writereg(sc->bge_dev, 1, 0x18, 0x02);
2244 #endif
2245 		DELAY(40); DELAY(40); DELAY(40);
2246 		DELAY(10000);	/* above not quite adequate on 5700 */
2247 		return 0;
2248 	}
2249 
2250 
2251 	/*
2252 	 * Entering ACPI power states D1-D3 is achieved by wiggling
2253 	 * GMII gpio pins. Example code assumes all hardware vendors
2254 	 * followed Broadom's sample pcb layout. Until we verify that
2255 	 * for all supported OEM cards, states D1-D3 are  unsupported.
2256 	 */
2257 	aprint_error_dev(sc->bge_dev,
2258 	    "power state %d unimplemented; check GPIO pins\n",
2259 	    powerlevel);
2260 #endif
2261 	return EOPNOTSUPP;
2262 }
2263 
2264 
2265 /*
2266  * Probe for a Broadcom chip. Check the PCI vendor and device IDs
2267  * against our list and return its name if we find a match. Note
2268  * that since the Broadcom controller contains VPD support, we
2269  * can get the device name string from the controller itself instead
2270  * of the compiled-in string. This is a little slow, but it guarantees
2271  * we'll always announce the right product name.
2272  */
2273 static int
2274 bge_probe(device_t parent, cfdata_t match, void *aux)
2275 {
2276 	struct pci_attach_args *pa = (struct pci_attach_args *)aux;
2277 
2278 	if (bge_lookup(pa) != NULL)
2279 		return (1);
2280 
2281 	return (0);
2282 }
2283 
2284 static void
2285 bge_attach(device_t parent, device_t self, void *aux)
2286 {
2287 	struct bge_softc	*sc = device_private(self);
2288 	struct pci_attach_args	*pa = aux;
2289 	const struct bge_product *bp;
2290 	const struct bge_revision *br;
2291 	pci_chipset_tag_t	pc;
2292 	pci_intr_handle_t	ih;
2293 	const char		*intrstr = NULL;
2294 	bus_dma_segment_t	seg;
2295 	int			rseg;
2296 	u_int32_t		hwcfg = 0;
2297 	u_int32_t		command;
2298 	struct ifnet		*ifp;
2299 	void *			kva;
2300 	u_char			eaddr[ETHER_ADDR_LEN];
2301 	pcireg_t		memtype;
2302 	bus_addr_t		memaddr;
2303 	bus_size_t		memsize;
2304 	u_int32_t		pm_ctl;
2305 
2306 	bp = bge_lookup(pa);
2307 	KASSERT(bp != NULL);
2308 
2309 	sc->sc_pc = pa->pa_pc;
2310 	sc->sc_pcitag = pa->pa_tag;
2311 	sc->bge_dev = self;
2312 
2313 	aprint_naive(": Ethernet controller\n");
2314 	aprint_normal(": %s\n", bp->bp_name);
2315 
2316 	/*
2317 	 * Map control/status registers.
2318 	 */
2319 	DPRINTFN(5, ("Map control/status regs\n"));
2320 	pc = sc->sc_pc;
2321 	command = pci_conf_read(pc, sc->sc_pcitag, PCI_COMMAND_STATUS_REG);
2322 	command |= PCI_COMMAND_MEM_ENABLE | PCI_COMMAND_MASTER_ENABLE;
2323 	pci_conf_write(pc, sc->sc_pcitag, PCI_COMMAND_STATUS_REG, command);
2324 	command = pci_conf_read(pc, sc->sc_pcitag, PCI_COMMAND_STATUS_REG);
2325 
2326 	if (!(command & PCI_COMMAND_MEM_ENABLE)) {
2327 		aprint_error_dev(sc->bge_dev,
2328 		    "failed to enable memory mapping!\n");
2329 		return;
2330 	}
2331 
2332 	DPRINTFN(5, ("pci_mem_find\n"));
2333 	memtype = pci_mapreg_type(sc->sc_pc, sc->sc_pcitag, BGE_PCI_BAR0);
2334  	switch (memtype) {
2335 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_32BIT:
2336 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_64BIT:
2337 		if (pci_mapreg_map(pa, BGE_PCI_BAR0,
2338 		    memtype, 0, &sc->bge_btag, &sc->bge_bhandle,
2339 		    &memaddr, &memsize) == 0)
2340 			break;
2341 	default:
2342 		aprint_error_dev(sc->bge_dev, "can't find mem space\n");
2343 		return;
2344 	}
2345 
2346 	DPRINTFN(5, ("pci_intr_map\n"));
2347 	if (pci_intr_map(pa, &ih)) {
2348 		aprint_error_dev(sc->bge_dev, "couldn't map interrupt\n");
2349 		return;
2350 	}
2351 
2352 	DPRINTFN(5, ("pci_intr_string\n"));
2353 	intrstr = pci_intr_string(pc, ih);
2354 
2355 	DPRINTFN(5, ("pci_intr_establish\n"));
2356 	sc->bge_intrhand = pci_intr_establish(pc, ih, IPL_NET, bge_intr, sc);
2357 
2358 	if (sc->bge_intrhand == NULL) {
2359 		aprint_error_dev(sc->bge_dev,
2360 		    "couldn't establish interrupt%s%s\n",
2361 		    intrstr ? " at " : "", intrstr ? intrstr : "");
2362 		return;
2363 	}
2364 	aprint_normal_dev(sc->bge_dev, "interrupting at %s\n", intrstr);
2365 
2366 	/*
2367 	 * Kludge for 5700 Bx bug: a hardware bug (PCIX byte enable?)
2368 	 * can clobber the chip's PCI config-space power control registers,
2369 	 * leaving the card in D3 powersave state.
2370 	 * We do not have memory-mapped registers in this state,
2371 	 * so force device into D0 state before starting initialization.
2372 	 */
2373 	pm_ctl = pci_conf_read(pc, sc->sc_pcitag, BGE_PCI_PWRMGMT_CMD);
2374 	pm_ctl &= ~(PCI_PWR_D0|PCI_PWR_D1|PCI_PWR_D2|PCI_PWR_D3);
2375 	pm_ctl |= (1 << 8) | PCI_PWR_D0 ; /* D0 state */
2376 	pci_conf_write(pc, sc->sc_pcitag, BGE_PCI_PWRMGMT_CMD, pm_ctl);
2377 	DELAY(1000);	/* 27 usec is allegedly sufficent */
2378 
2379 	/*
2380 	 * Save ASIC rev.  Look up any quirks associated with this
2381 	 * ASIC.
2382 	 */
2383 	sc->bge_chipid =
2384 	    pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MISC_CTL) &
2385 	    BGE_PCIMISCCTL_ASICREV;
2386 
2387 	/*
2388 	 * Detect PCI-Express devices
2389 	 * XXX: guessed from Linux/FreeBSD; no documentation
2390 	 */
2391 	if (pci_get_capability(sc->sc_pc, sc->sc_pcitag, PCI_CAP_PCIEXPRESS,
2392 	        NULL, NULL) != 0)
2393 		sc->bge_flags |= BGE_PCIE;
2394 
2395 	/*
2396 	 * PCI-X check.
2397 	 */
2398 	if ((pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_PCISTATE) &
2399 		BGE_PCISTATE_PCI_BUSMODE) == 0)
2400 		sc->bge_flags |= BGE_PCIX;
2401 
2402 	/* Try to reset the chip. */
2403 	DPRINTFN(5, ("bge_reset\n"));
2404 	bge_reset(sc);
2405 
2406 	if (bge_chipinit(sc)) {
2407 		aprint_error_dev(sc->bge_dev, "chip initialization failed\n");
2408 		bge_release_resources(sc);
2409 		return;
2410 	}
2411 
2412 	/*
2413 	 * Get station address from the EEPROM.
2414 	 */
2415 	if (bge_get_eaddr(sc, eaddr)) {
2416 		aprint_error_dev(sc->bge_dev,
2417 		"failed to reade station address\n");
2418 		bge_release_resources(sc);
2419 		return;
2420 	}
2421 
2422 	br = bge_lookup_rev(sc->bge_chipid);
2423 
2424 	if (br == NULL) {
2425 		aprint_normal_dev(sc->bge_dev, "unknown ASIC (0x%04x)",
2426 		    sc->bge_chipid >> 16);
2427 	} else {
2428 		aprint_normal_dev(sc->bge_dev, "ASIC %s (0x%04x)",
2429 		    br->br_name, sc->bge_chipid >> 16);
2430 	}
2431 	aprint_normal(", Ethernet address %s\n", ether_sprintf(eaddr));
2432 
2433 	/* Allocate the general information block and ring buffers. */
2434 	if (pci_dma64_available(pa))
2435 		sc->bge_dmatag = pa->pa_dmat64;
2436 	else
2437 		sc->bge_dmatag = pa->pa_dmat;
2438 	DPRINTFN(5, ("bus_dmamem_alloc\n"));
2439 	if (bus_dmamem_alloc(sc->bge_dmatag, sizeof(struct bge_ring_data),
2440 			     PAGE_SIZE, 0, &seg, 1, &rseg, BUS_DMA_NOWAIT)) {
2441 		aprint_error_dev(sc->bge_dev, "can't alloc rx buffers\n");
2442 		return;
2443 	}
2444 	DPRINTFN(5, ("bus_dmamem_map\n"));
2445 	if (bus_dmamem_map(sc->bge_dmatag, &seg, rseg,
2446 			   sizeof(struct bge_ring_data), &kva,
2447 			   BUS_DMA_NOWAIT)) {
2448 		aprint_error_dev(sc->bge_dev,
2449 		    "can't map DMA buffers (%zu bytes)\n",
2450 		    sizeof(struct bge_ring_data));
2451 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2452 		return;
2453 	}
2454 	DPRINTFN(5, ("bus_dmamem_create\n"));
2455 	if (bus_dmamap_create(sc->bge_dmatag, sizeof(struct bge_ring_data), 1,
2456 	    sizeof(struct bge_ring_data), 0,
2457 	    BUS_DMA_NOWAIT, &sc->bge_ring_map)) {
2458 		aprint_error_dev(sc->bge_dev, "can't create DMA map\n");
2459 		bus_dmamem_unmap(sc->bge_dmatag, kva,
2460 				 sizeof(struct bge_ring_data));
2461 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2462 		return;
2463 	}
2464 	DPRINTFN(5, ("bus_dmamem_load\n"));
2465 	if (bus_dmamap_load(sc->bge_dmatag, sc->bge_ring_map, kva,
2466 			    sizeof(struct bge_ring_data), NULL,
2467 			    BUS_DMA_NOWAIT)) {
2468 		bus_dmamap_destroy(sc->bge_dmatag, sc->bge_ring_map);
2469 		bus_dmamem_unmap(sc->bge_dmatag, kva,
2470 				 sizeof(struct bge_ring_data));
2471 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2472 		return;
2473 	}
2474 
2475 	DPRINTFN(5, ("bzero\n"));
2476 	sc->bge_rdata = (struct bge_ring_data *)kva;
2477 
2478 	memset(sc->bge_rdata, 0, sizeof(struct bge_ring_data));
2479 
2480 	/* Try to allocate memory for jumbo buffers. */
2481 	if (!(BGE_IS_5705_OR_BEYOND(sc))) {
2482 		if (bge_alloc_jumbo_mem(sc)) {
2483 			aprint_error_dev(sc->bge_dev,
2484 			    "jumbo buffer allocation failed\n");
2485 		} else
2486 			sc->ethercom.ec_capabilities |= ETHERCAP_JUMBO_MTU;
2487 	}
2488 
2489 	/* Set default tuneable values. */
2490 	sc->bge_stat_ticks = BGE_TICKS_PER_SEC;
2491 	sc->bge_rx_coal_ticks = 150;
2492 	sc->bge_rx_max_coal_bds = 64;
2493 #ifdef ORIG_WPAUL_VALUES
2494 	sc->bge_tx_coal_ticks = 150;
2495 	sc->bge_tx_max_coal_bds = 128;
2496 #else
2497 	sc->bge_tx_coal_ticks = 300;
2498 	sc->bge_tx_max_coal_bds = 400;
2499 #endif
2500 	if (BGE_IS_5705_OR_BEYOND(sc)) {
2501 		sc->bge_tx_coal_ticks = (12 * 5);
2502 		sc->bge_tx_max_coal_bds = (12 * 5);
2503 			aprint_verbose_dev(sc->bge_dev,
2504 			    "setting short Tx thresholds\n");
2505 	}
2506 
2507 	/* Set up ifnet structure */
2508 	ifp = &sc->ethercom.ec_if;
2509 	ifp->if_softc = sc;
2510 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
2511 	ifp->if_ioctl = bge_ioctl;
2512 	ifp->if_stop = bge_stop;
2513 	ifp->if_start = bge_start;
2514 	ifp->if_init = bge_init;
2515 	ifp->if_watchdog = bge_watchdog;
2516 	IFQ_SET_MAXLEN(&ifp->if_snd, max(BGE_TX_RING_CNT - 1, IFQ_MAXLEN));
2517 	IFQ_SET_READY(&ifp->if_snd);
2518 	DPRINTFN(5, ("strcpy if_xname\n"));
2519 	strcpy(ifp->if_xname, device_xname(sc->bge_dev));
2520 
2521 	if (sc->bge_chipid != BGE_CHIPID_BCM5700_B0)
2522 		sc->ethercom.ec_if.if_capabilities |=
2523 		    IFCAP_CSUM_IPv4_Tx | IFCAP_CSUM_IPv4_Rx |
2524 		    IFCAP_CSUM_TCPv4_Tx | IFCAP_CSUM_TCPv4_Rx |
2525 		    IFCAP_CSUM_UDPv4_Tx | IFCAP_CSUM_UDPv4_Rx;
2526 	sc->ethercom.ec_capabilities |=
2527 	    ETHERCAP_VLAN_HWTAGGING | ETHERCAP_VLAN_MTU;
2528 
2529 	if (sc->bge_flags & BGE_PCIE)
2530 		sc->ethercom.ec_if.if_capabilities |= IFCAP_TSOv4;
2531 
2532 	/*
2533 	 * Do MII setup.
2534 	 */
2535 	DPRINTFN(5, ("mii setup\n"));
2536 	sc->bge_mii.mii_ifp = ifp;
2537 	sc->bge_mii.mii_readreg = bge_miibus_readreg;
2538 	sc->bge_mii.mii_writereg = bge_miibus_writereg;
2539 	sc->bge_mii.mii_statchg = bge_miibus_statchg;
2540 
2541 	/*
2542 	 * Figure out what sort of media we have by checking the
2543 	 * hardware config word in the first 32k of NIC internal memory,
2544 	 * or fall back to the config word in the EEPROM. Note: on some BCM5700
2545 	 * cards, this value appears to be unset. If that's the
2546 	 * case, we have to rely on identifying the NIC by its PCI
2547 	 * subsystem ID, as we do below for the SysKonnect SK-9D41.
2548 	 */
2549 	if (bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_SIG) == BGE_MAGIC_NUMBER) {
2550 		hwcfg = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_NICCFG);
2551 	} else {
2552 		bge_read_eeprom(sc, (void *)&hwcfg,
2553 		    BGE_EE_HWCFG_OFFSET, sizeof(hwcfg));
2554 		hwcfg = be32toh(hwcfg);
2555 	}
2556 	if ((hwcfg & BGE_HWCFG_MEDIA) == BGE_MEDIA_FIBER)
2557 		sc->bge_flags |= BGE_PHY_FIBER_TBI;
2558 
2559 	/* The SysKonnect SK-9D41 is a 1000baseSX card. */
2560 	if ((pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_SUBSYS) >> 16) ==
2561 	    SK_SUBSYSID_9D41)
2562 		sc->bge_flags |= BGE_PHY_FIBER_TBI;
2563 
2564 	if (sc->bge_flags & BGE_PHY_FIBER_TBI) {
2565 		ifmedia_init(&sc->bge_ifmedia, IFM_IMASK, bge_ifmedia_upd,
2566 		    bge_ifmedia_sts);
2567 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX, 0, NULL);
2568 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX|IFM_FDX,
2569 			    0, NULL);
2570 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO, 0, NULL);
2571 		ifmedia_set(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO);
2572 		/* Pretend the user requested this setting */
2573 		sc->bge_ifmedia.ifm_media =
2574 			sc->bge_ifmedia.ifm_cur->ifm_media;
2575 	} else {
2576 		/*
2577 		 * Do transceiver setup.
2578 		 */
2579 		ifmedia_init(&sc->bge_mii.mii_media, 0, bge_ifmedia_upd,
2580 			     bge_ifmedia_sts);
2581 		mii_attach(sc->bge_dev, &sc->bge_mii, 0xffffffff,
2582 			   MII_PHY_ANY, MII_OFFSET_ANY,
2583 			   MIIF_FORCEANEG|MIIF_DOPAUSE);
2584 
2585 		if (LIST_EMPTY(&sc->bge_mii.mii_phys)) {
2586 			aprint_error_dev(sc->bge_dev, "no PHY found!\n");
2587 			ifmedia_add(&sc->bge_mii.mii_media,
2588 				    IFM_ETHER|IFM_MANUAL, 0, NULL);
2589 			ifmedia_set(&sc->bge_mii.mii_media,
2590 				    IFM_ETHER|IFM_MANUAL);
2591 		} else
2592 			ifmedia_set(&sc->bge_mii.mii_media,
2593 				    IFM_ETHER|IFM_AUTO);
2594 	}
2595 
2596 	/*
2597 	 * When using the BCM5701 in PCI-X mode, data corruption has
2598 	 * been observed in the first few bytes of some received packets.
2599 	 * Aligning the packet buffer in memory eliminates the corruption.
2600 	 * Unfortunately, this misaligns the packet payloads.  On platforms
2601 	 * which do not support unaligned accesses, we will realign the
2602 	 * payloads by copying the received packets.
2603 	 */
2604 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5701 &&
2605 		sc->bge_flags & BGE_PCIX)
2606 		sc->bge_flags |= BGE_RX_ALIGNBUG;
2607 
2608 	/*
2609 	 * Call MI attach routine.
2610 	 */
2611 	DPRINTFN(5, ("if_attach\n"));
2612 	if_attach(ifp);
2613 	DPRINTFN(5, ("ether_ifattach\n"));
2614 	ether_ifattach(ifp, eaddr);
2615 #if NRND > 0
2616 	rnd_attach_source(&sc->rnd_source, device_xname(sc->bge_dev),
2617 		RND_TYPE_NET, 0);
2618 #endif
2619 #ifdef BGE_EVENT_COUNTERS
2620 	/*
2621 	 * Attach event counters.
2622 	 */
2623 	evcnt_attach_dynamic(&sc->bge_ev_intr, EVCNT_TYPE_INTR,
2624 	    NULL, device_xname(sc->bge_dev), "intr");
2625 	evcnt_attach_dynamic(&sc->bge_ev_tx_xoff, EVCNT_TYPE_MISC,
2626 	    NULL, device_xname(sc->bge_dev), "tx_xoff");
2627 	evcnt_attach_dynamic(&sc->bge_ev_tx_xon, EVCNT_TYPE_MISC,
2628 	    NULL, device_xname(sc->bge_dev), "tx_xon");
2629 	evcnt_attach_dynamic(&sc->bge_ev_rx_xoff, EVCNT_TYPE_MISC,
2630 	    NULL, device_xname(sc->bge_dev), "rx_xoff");
2631 	evcnt_attach_dynamic(&sc->bge_ev_rx_xon, EVCNT_TYPE_MISC,
2632 	    NULL, device_xname(sc->bge_dev), "rx_xon");
2633 	evcnt_attach_dynamic(&sc->bge_ev_rx_macctl, EVCNT_TYPE_MISC,
2634 	    NULL, device_xname(sc->bge_dev), "rx_macctl");
2635 	evcnt_attach_dynamic(&sc->bge_ev_xoffentered, EVCNT_TYPE_MISC,
2636 	    NULL, device_xname(sc->bge_dev), "xoffentered");
2637 #endif /* BGE_EVENT_COUNTERS */
2638 	DPRINTFN(5, ("callout_init\n"));
2639 	callout_init(&sc->bge_timeout, 0);
2640 
2641 	if (!pmf_device_register(self, NULL, NULL))
2642 		aprint_error_dev(self, "couldn't establish power handler\n");
2643 	else
2644 		pmf_class_network_register(self, ifp);
2645 }
2646 
2647 static void
2648 bge_release_resources(struct bge_softc *sc)
2649 {
2650 	if (sc->bge_vpd_prodname != NULL)
2651 		free(sc->bge_vpd_prodname, M_DEVBUF);
2652 
2653 	if (sc->bge_vpd_readonly != NULL)
2654 		free(sc->bge_vpd_readonly, M_DEVBUF);
2655 }
2656 
2657 static void
2658 bge_reset(struct bge_softc *sc)
2659 {
2660 	u_int32_t cachesize, command, pcistate, new_pcistate;
2661 	int i, val;
2662 	void (*write_op)(struct bge_softc *, int, int);
2663 
2664 	if (BGE_IS_5750_OR_BEYOND(sc) && !BGE_IS_5714_FAMILY(sc) &&
2665 	    (BGE_ASICREV(sc->bge_chipid) != BGE_ASICREV_BCM5906)) {
2666 	    	if (sc->bge_flags & BGE_PCIE) {
2667 			write_op = bge_writemem_direct;
2668 		} else {
2669 			write_op = bge_writemem_ind;
2670 		}
2671 	} else {
2672 		write_op = bge_writereg_ind;
2673 	}
2674 
2675 
2676 	/* Save some important PCI state. */
2677 	cachesize = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CACHESZ);
2678 	command = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CMD);
2679 	pcistate = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_PCISTATE);
2680 
2681 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MISC_CTL,
2682 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
2683 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
2684 
2685 	/*
2686 	 * Disable the firmware fastboot feature on 5752 ASIC
2687 	 * to avoid firmware timeout.
2688 	 */
2689 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5752 ||
2690 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5755 ||
2691 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5787)
2692 		CSR_WRITE_4(sc, BGE_FASTBOOT_PC, 0);
2693 
2694 	val = BGE_MISCCFG_RESET_CORE_CLOCKS | (65<<1);
2695 	/*
2696 	 * XXX: from FreeBSD/Linux; no documentation
2697 	 */
2698 	if (sc->bge_flags & BGE_PCIE) {
2699 		if (CSR_READ_4(sc, BGE_PCIE_CTL1) == 0x60)
2700 			/* PCI Express 1.0 system */
2701 			CSR_WRITE_4(sc, BGE_PCIE_CTL1, 0x20);
2702 		if (sc->bge_chipid != BGE_CHIPID_BCM5750_A0) {
2703 			/*
2704 			 * Prevent PCI Express link training
2705 			 * during global reset.
2706 			 */
2707 			CSR_WRITE_4(sc, BGE_MISC_CFG, 1 << 29);
2708 			val |= (1<<29);
2709 		}
2710 	}
2711 
2712 	/* Issue global reset */
2713 	write_op(sc, BGE_MISC_CFG, val);
2714 
2715 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5906) {
2716 		i = CSR_READ_4(sc, BGE_VCPU_STATUS);
2717 		CSR_WRITE_4(sc, BGE_VCPU_STATUS,
2718 		    i | BGE_VCPU_STATUS_DRV_RESET);
2719 		i = CSR_READ_4(sc, BGE_VCPU_EXT_CTRL);
2720 		CSR_WRITE_4(sc, BGE_VCPU_EXT_CTRL,
2721 		    i & ~BGE_VCPU_EXT_CTRL_HALT_CPU);
2722 	}
2723 
2724 
2725 
2726 	DELAY(1000);
2727 
2728 	/*
2729 	 * XXX: from FreeBSD/Linux; no documentation
2730 	 */
2731 	if (sc->bge_flags & BGE_PCIE) {
2732 		if (sc->bge_chipid == BGE_CHIPID_BCM5750_A0) {
2733 			pcireg_t reg;
2734 
2735 			DELAY(500000);
2736 			/* XXX: Magic Numbers */
2737 			reg = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_UNKNOWN0);
2738 			pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_UNKNOWN0,
2739 			    reg | (1 << 15));
2740 		}
2741 		/*
2742 		 * XXX: Magic Numbers.
2743 		 * Sets maximal PCI-e payload and clears any PCI-e errors.
2744 		 * Should be replaced with references to PCI config-space
2745 		 * capability block for PCI-Express.
2746 		 */
2747 		pci_conf_write(sc->sc_pc, sc->sc_pcitag,
2748 		    BGE_PCI_CONF_DEV_CTRL, 0xf5000);
2749 
2750 	}
2751 
2752 	/* Reset some of the PCI state that got zapped by reset */
2753 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MISC_CTL,
2754 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
2755 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
2756 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CMD, command);
2757 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CACHESZ, cachesize);
2758 	write_op(sc, BGE_MISC_CFG, (65 << 1));
2759 
2760 	/* Enable memory arbiter. */
2761 	{
2762 		uint32_t marbmode = 0;
2763 		if (BGE_IS_5714_FAMILY(sc)) {
2764 			marbmode = CSR_READ_4(sc, BGE_MARB_MODE);
2765 		}
2766  		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE | marbmode);
2767 	}
2768 
2769 
2770 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5906) {
2771 		for (i = 0; i < BGE_TIMEOUT; i++) {
2772 			val = CSR_READ_4(sc, BGE_VCPU_STATUS);
2773 			if (val & BGE_VCPU_STATUS_INIT_DONE)
2774 				break;
2775 			DELAY(100);
2776 		}
2777 		if (i == BGE_TIMEOUT) {
2778 			aprint_error_dev(sc->bge_dev, "reset timed out\n");
2779 			return;
2780 		}
2781 	} else {
2782 		/*
2783 		 * Write the magic number to the firmware mailbox at 0xb50
2784 		 * so that the driver can synchronize with the firmware.
2785 		 */
2786 		bge_writemem_ind(sc, BGE_SOFTWARE_GENCOMM, BGE_MAGIC_NUMBER);
2787 
2788 		/*
2789 		 * Poll the value location we just wrote until
2790 		 * we see the 1's complement of the magic number.
2791 		 * This indicates that the firmware initialization
2792 		 * is complete.
2793 		 */
2794 		for (i = 0; i < BGE_TIMEOUT; i++) {
2795 			val = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM);
2796 			if (val == ~BGE_MAGIC_NUMBER)
2797 				break;
2798 			DELAY(1000);
2799 		}
2800 
2801 		if (i >= BGE_TIMEOUT) {
2802 			aprint_error_dev(sc->bge_dev,
2803 			    "firmware handshake timed out, val = %x\n", val);
2804 			/*
2805 			 * XXX: occasionally fired on bcm5721, but without
2806 			 * apparent harm.  For now, keep going if we timeout
2807 			 * against PCI-E devices.
2808 			 */
2809 			if ((sc->bge_flags & BGE_PCIE) == 0)
2810 				  return;
2811 		}
2812 	}
2813 
2814 	/*
2815 	 * XXX Wait for the value of the PCISTATE register to
2816 	 * return to its original pre-reset state. This is a
2817 	 * fairly good indicator of reset completion. If we don't
2818 	 * wait for the reset to fully complete, trying to read
2819 	 * from the device's non-PCI registers may yield garbage
2820 	 * results.
2821 	 */
2822 	for (i = 0; i < 10000; i++) {
2823 		new_pcistate = pci_conf_read(sc->sc_pc, sc->sc_pcitag,
2824 		    BGE_PCI_PCISTATE);
2825 		if ((new_pcistate & ~BGE_PCISTATE_RESERVED) ==
2826 		    (pcistate & ~BGE_PCISTATE_RESERVED))
2827 			break;
2828 		DELAY(10);
2829 	}
2830 	if ((new_pcistate & ~BGE_PCISTATE_RESERVED) !=
2831 	    (pcistate & ~BGE_PCISTATE_RESERVED)) {
2832 		aprint_error_dev(sc->bge_dev, "pcistate failed to revert\n");
2833 	}
2834 
2835 	/* XXX: from FreeBSD/Linux; no documentation */
2836 	if (sc->bge_flags & BGE_PCIE &&
2837 	    sc->bge_chipid != BGE_CHIPID_BCM5750_A0)
2838 		CSR_WRITE_4(sc, BGE_PCIE_CTL0, CSR_READ_4(sc, BGE_PCIE_CTL0) | (1<<25));
2839 
2840 	/* Enable memory arbiter. */
2841 	/* XXX why do this twice? */
2842 	{
2843 		uint32_t marbmode = 0;
2844 		if (BGE_IS_5714_FAMILY(sc)) {
2845 			marbmode = CSR_READ_4(sc, BGE_MARB_MODE);
2846 		}
2847  		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE | marbmode);
2848 	}
2849 
2850 	/* Fix up byte swapping */
2851 	CSR_WRITE_4(sc, BGE_MODE_CTL, BGE_DMA_SWAP_OPTIONS);
2852 
2853 	CSR_WRITE_4(sc, BGE_MAC_MODE, 0);
2854 
2855 	DELAY(10000);
2856 }
2857 
2858 /*
2859  * Frame reception handling. This is called if there's a frame
2860  * on the receive return list.
2861  *
2862  * Note: we have to be able to handle two possibilities here:
2863  * 1) the frame is from the jumbo recieve ring
2864  * 2) the frame is from the standard receive ring
2865  */
2866 
2867 static void
2868 bge_rxeof(struct bge_softc *sc)
2869 {
2870 	struct ifnet *ifp;
2871 	int stdcnt = 0, jumbocnt = 0;
2872 	bus_dmamap_t dmamap;
2873 	bus_addr_t offset, toff;
2874 	bus_size_t tlen;
2875 	int tosync;
2876 
2877 	ifp = &sc->ethercom.ec_if;
2878 
2879 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2880 	    offsetof(struct bge_ring_data, bge_status_block),
2881 	    sizeof (struct bge_status_block),
2882 	    BUS_DMASYNC_POSTREAD);
2883 
2884 	offset = offsetof(struct bge_ring_data, bge_rx_return_ring);
2885 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx -
2886 	    sc->bge_rx_saved_considx;
2887 
2888 #if NRND > 0
2889 	if (tosync != 0 && RND_ENABLED(&sc->rnd_source))
2890 		rnd_add_uint32(&sc->rnd_source, tosync);
2891 #endif
2892 
2893 	toff = offset + (sc->bge_rx_saved_considx * sizeof (struct bge_rx_bd));
2894 
2895 	if (tosync < 0) {
2896 		tlen = (sc->bge_return_ring_cnt - sc->bge_rx_saved_considx) *
2897 		    sizeof (struct bge_rx_bd);
2898 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2899 		    toff, tlen, BUS_DMASYNC_POSTREAD);
2900 		tosync = -tosync;
2901 	}
2902 
2903 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2904 	    offset, tosync * sizeof (struct bge_rx_bd),
2905 	    BUS_DMASYNC_POSTREAD);
2906 
2907 	while(sc->bge_rx_saved_considx !=
2908 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx) {
2909 		struct bge_rx_bd	*cur_rx;
2910 		u_int32_t		rxidx;
2911 		struct mbuf		*m = NULL;
2912 
2913 		cur_rx = &sc->bge_rdata->
2914 			bge_rx_return_ring[sc->bge_rx_saved_considx];
2915 
2916 		rxidx = cur_rx->bge_idx;
2917 		BGE_INC(sc->bge_rx_saved_considx, sc->bge_return_ring_cnt);
2918 
2919 		if (cur_rx->bge_flags & BGE_RXBDFLAG_JUMBO_RING) {
2920 			BGE_INC(sc->bge_jumbo, BGE_JUMBO_RX_RING_CNT);
2921 			m = sc->bge_cdata.bge_rx_jumbo_chain[rxidx];
2922 			sc->bge_cdata.bge_rx_jumbo_chain[rxidx] = NULL;
2923 			jumbocnt++;
2924 			bus_dmamap_sync(sc->bge_dmatag,
2925 			    sc->bge_cdata.bge_rx_jumbo_map,
2926 			    mtod(m, char *) - (char *)sc->bge_cdata.bge_jumbo_buf,
2927 			    BGE_JLEN, BUS_DMASYNC_POSTREAD);
2928 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
2929 				ifp->if_ierrors++;
2930 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
2931 				continue;
2932 			}
2933 			if (bge_newbuf_jumbo(sc, sc->bge_jumbo,
2934 					     NULL)== ENOBUFS) {
2935 				ifp->if_ierrors++;
2936 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
2937 				continue;
2938 			}
2939 		} else {
2940 			BGE_INC(sc->bge_std, BGE_STD_RX_RING_CNT);
2941 			m = sc->bge_cdata.bge_rx_std_chain[rxidx];
2942 
2943 			sc->bge_cdata.bge_rx_std_chain[rxidx] = NULL;
2944 			stdcnt++;
2945 			dmamap = sc->bge_cdata.bge_rx_std_map[rxidx];
2946 			sc->bge_cdata.bge_rx_std_map[rxidx] = 0;
2947 			bus_dmamap_sync(sc->bge_dmatag, dmamap, 0,
2948 			    dmamap->dm_mapsize, BUS_DMASYNC_POSTREAD);
2949 			bus_dmamap_unload(sc->bge_dmatag, dmamap);
2950 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
2951 				ifp->if_ierrors++;
2952 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
2953 				continue;
2954 			}
2955 			if (bge_newbuf_std(sc, sc->bge_std,
2956 			    NULL, dmamap) == ENOBUFS) {
2957 				ifp->if_ierrors++;
2958 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
2959 				continue;
2960 			}
2961 		}
2962 
2963 		ifp->if_ipackets++;
2964 #ifndef __NO_STRICT_ALIGNMENT
2965                 /*
2966                  * XXX: if the 5701 PCIX-Rx-DMA workaround is in effect,
2967                  * the Rx buffer has the layer-2 header unaligned.
2968                  * If our CPU requires alignment, re-align by copying.
2969                  */
2970 		if (sc->bge_flags & BGE_RX_ALIGNBUG) {
2971 			memmove(mtod(m, char *) + ETHER_ALIGN, m->m_data,
2972                                 cur_rx->bge_len);
2973 			m->m_data += ETHER_ALIGN;
2974 		}
2975 #endif
2976 
2977 		m->m_pkthdr.len = m->m_len = cur_rx->bge_len - ETHER_CRC_LEN;
2978 		m->m_pkthdr.rcvif = ifp;
2979 
2980 #if NBPFILTER > 0
2981 		/*
2982 		 * Handle BPF listeners. Let the BPF user see the packet.
2983 		 */
2984 		if (ifp->if_bpf)
2985 			bpf_mtap(ifp->if_bpf, m);
2986 #endif
2987 
2988 		m->m_pkthdr.csum_flags = M_CSUM_IPv4;
2989 
2990 		if ((cur_rx->bge_ip_csum ^ 0xffff) != 0)
2991 			m->m_pkthdr.csum_flags |= M_CSUM_IPv4_BAD;
2992 		/*
2993 		 * Rx transport checksum-offload may also
2994 		 * have bugs with packets which, when transmitted,
2995 		 * were `runts' requiring padding.
2996 		 */
2997 		if (cur_rx->bge_flags & BGE_RXBDFLAG_TCP_UDP_CSUM &&
2998 		    (/* (sc->_bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||*/
2999 		     m->m_pkthdr.len >= ETHER_MIN_NOPAD)) {
3000 			m->m_pkthdr.csum_data =
3001 			    cur_rx->bge_tcp_udp_csum;
3002 			m->m_pkthdr.csum_flags |=
3003 			    (M_CSUM_TCPv4|M_CSUM_UDPv4|
3004 			     M_CSUM_DATA|M_CSUM_NO_PSEUDOHDR);
3005 		}
3006 
3007 		/*
3008 		 * If we received a packet with a vlan tag, pass it
3009 		 * to vlan_input() instead of ether_input().
3010 		 */
3011 		if (cur_rx->bge_flags & BGE_RXBDFLAG_VLAN_TAG) {
3012 			VLAN_INPUT_TAG(ifp, m, cur_rx->bge_vlan_tag, continue);
3013 		}
3014 
3015 		(*ifp->if_input)(ifp, m);
3016 	}
3017 
3018 	bge_writembx(sc, BGE_MBX_RX_CONS0_LO, sc->bge_rx_saved_considx);
3019 	if (stdcnt)
3020 		bge_writembx(sc, BGE_MBX_RX_STD_PROD_LO, sc->bge_std);
3021 	if (jumbocnt)
3022 		bge_writembx(sc, BGE_MBX_RX_JUMBO_PROD_LO, sc->bge_jumbo);
3023 }
3024 
3025 static void
3026 bge_txeof(struct bge_softc *sc)
3027 {
3028 	struct bge_tx_bd *cur_tx = NULL;
3029 	struct ifnet *ifp;
3030 	struct txdmamap_pool_entry *dma;
3031 	bus_addr_t offset, toff;
3032 	bus_size_t tlen;
3033 	int tosync;
3034 	struct mbuf *m;
3035 
3036 	ifp = &sc->ethercom.ec_if;
3037 
3038 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
3039 	    offsetof(struct bge_ring_data, bge_status_block),
3040 	    sizeof (struct bge_status_block),
3041 	    BUS_DMASYNC_POSTREAD);
3042 
3043 	offset = offsetof(struct bge_ring_data, bge_tx_ring);
3044 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx -
3045 	    sc->bge_tx_saved_considx;
3046 
3047 #if NRND > 0
3048 	if (tosync != 0 && RND_ENABLED(&sc->rnd_source))
3049 		rnd_add_uint32(&sc->rnd_source, tosync);
3050 #endif
3051 
3052 	toff = offset + (sc->bge_tx_saved_considx * sizeof (struct bge_tx_bd));
3053 
3054 	if (tosync < 0) {
3055 		tlen = (BGE_TX_RING_CNT - sc->bge_tx_saved_considx) *
3056 		    sizeof (struct bge_tx_bd);
3057 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
3058 		    toff, tlen, BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
3059 		tosync = -tosync;
3060 	}
3061 
3062 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
3063 	    offset, tosync * sizeof (struct bge_tx_bd),
3064 	    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
3065 
3066 	/*
3067 	 * Go through our tx ring and free mbufs for those
3068 	 * frames that have been sent.
3069 	 */
3070 	while (sc->bge_tx_saved_considx !=
3071 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx) {
3072 		u_int32_t		idx = 0;
3073 
3074 		idx = sc->bge_tx_saved_considx;
3075 		cur_tx = &sc->bge_rdata->bge_tx_ring[idx];
3076 		if (cur_tx->bge_flags & BGE_TXBDFLAG_END)
3077 			ifp->if_opackets++;
3078 		m = sc->bge_cdata.bge_tx_chain[idx];
3079 		if (m != NULL) {
3080 			sc->bge_cdata.bge_tx_chain[idx] = NULL;
3081 			dma = sc->txdma[idx];
3082 			bus_dmamap_sync(sc->bge_dmatag, dma->dmamap, 0,
3083 			    dma->dmamap->dm_mapsize, BUS_DMASYNC_POSTWRITE);
3084 			bus_dmamap_unload(sc->bge_dmatag, dma->dmamap);
3085 			SLIST_INSERT_HEAD(&sc->txdma_list, dma, link);
3086 			sc->txdma[idx] = NULL;
3087 
3088 			m_freem(m);
3089 		}
3090 		sc->bge_txcnt--;
3091 		BGE_INC(sc->bge_tx_saved_considx, BGE_TX_RING_CNT);
3092 		ifp->if_timer = 0;
3093 	}
3094 
3095 	if (cur_tx != NULL)
3096 		ifp->if_flags &= ~IFF_OACTIVE;
3097 }
3098 
3099 static int
3100 bge_intr(void *xsc)
3101 {
3102 	struct bge_softc *sc;
3103 	struct ifnet *ifp;
3104 
3105 	sc = xsc;
3106 	ifp = &sc->ethercom.ec_if;
3107 
3108 	/*
3109 	 * Ascertain whether the interrupt is from this bge device.
3110 	 * Do the cheap test first.
3111 	 */
3112 	if ((sc->bge_rdata->bge_status_block.bge_status &
3113 	    BGE_STATFLAG_UPDATED) == 0) {
3114 		/*
3115 		 * Sometimes, the interrupt comes in before the
3116 		 * DMA update of the status block (performed prior
3117 		 * to the  interrupt itself) has completed.
3118 		 * In that case, do the (extremely expensive!)
3119 		 * PCI-config-space register read.
3120 		 */
3121 		uint32_t pcistate =
3122 		    pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_PCISTATE);
3123 
3124 		if (pcistate & BGE_PCISTATE_INTR_STATE)
3125 			return (0);
3126 
3127 	}
3128 	/*
3129 	 *  If we reach here, then the interrupt is for us.
3130 	 */
3131 
3132 	/* Ack interrupt and stop others from occuring. */
3133 	bge_writembx(sc, BGE_MBX_IRQ0_LO, 1);
3134 
3135 	BGE_EVCNT_INCR(sc->bge_ev_intr);
3136 
3137 	/*
3138 	 * Process link state changes.
3139 	 * Grrr. The link status word in the status block does
3140 	 * not work correctly on the BCM5700 rev AX and BX chips,
3141 	 * according to all available information. Hence, we have
3142 	 * to enable MII interrupts in order to properly obtain
3143 	 * async link changes. Unfortunately, this also means that
3144 	 * we have to read the MAC status register to detect link
3145 	 * changes, thereby adding an additional register access to
3146 	 * the interrupt handler.
3147 	 */
3148 
3149 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5700) {
3150 		u_int32_t		status;
3151 
3152 		status = CSR_READ_4(sc, BGE_MAC_STS);
3153 		if (status & BGE_MACSTAT_MI_INTERRUPT) {
3154 			sc->bge_link = 0;
3155 			callout_stop(&sc->bge_timeout);
3156 			bge_tick(sc);
3157 			/* Clear the interrupt */
3158 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
3159 			    BGE_EVTENB_MI_INTERRUPT);
3160 			bge_miibus_readreg(sc->bge_dev, 1, BRGPHY_MII_ISR);
3161 			bge_miibus_writereg(sc->bge_dev, 1, BRGPHY_MII_IMR,
3162 			    BRGPHY_INTRS);
3163 		}
3164 	} else {
3165 		u_int32_t		status;
3166 
3167 		status = CSR_READ_4(sc, BGE_MAC_STS);
3168 		if (status & BGE_MACSTAT_LINK_CHANGED) {
3169 			sc->bge_link = 0;
3170 			callout_stop(&sc->bge_timeout);
3171 			bge_tick(sc);
3172 			/* Clear the interrupt */
3173 			CSR_WRITE_4(sc, BGE_MAC_STS, BGE_MACSTAT_SYNC_CHANGED|
3174 			    BGE_MACSTAT_CFG_CHANGED|BGE_MACSTAT_MI_COMPLETE|
3175 			    BGE_MACSTAT_LINK_CHANGED);
3176 		}
3177 	}
3178 
3179 	if (ifp->if_flags & IFF_RUNNING) {
3180 		/* Check RX return ring producer/consumer */
3181 		bge_rxeof(sc);
3182 
3183 		/* Check TX ring producer/consumer */
3184 		bge_txeof(sc);
3185 	}
3186 
3187 	if (sc->bge_pending_rxintr_change) {
3188 		uint32_t rx_ticks = sc->bge_rx_coal_ticks;
3189 		uint32_t rx_bds = sc->bge_rx_max_coal_bds;
3190 		uint32_t junk;
3191 
3192 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, rx_ticks);
3193 		DELAY(10);
3194 		junk = CSR_READ_4(sc, BGE_HCC_RX_COAL_TICKS);
3195 
3196 		CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, rx_bds);
3197 		DELAY(10);
3198 		junk = CSR_READ_4(sc, BGE_HCC_RX_MAX_COAL_BDS);
3199 
3200 		sc->bge_pending_rxintr_change = 0;
3201 	}
3202 	bge_handle_events(sc);
3203 
3204 	/* Re-enable interrupts. */
3205 	bge_writembx(sc, BGE_MBX_IRQ0_LO, 0);
3206 
3207 	if (ifp->if_flags & IFF_RUNNING && !IFQ_IS_EMPTY(&ifp->if_snd))
3208 		bge_start(ifp);
3209 
3210 	return (1);
3211 }
3212 
3213 static void
3214 bge_tick(void *xsc)
3215 {
3216 	struct bge_softc *sc = xsc;
3217 	struct mii_data *mii = &sc->bge_mii;
3218 	int s;
3219 
3220 	s = splnet();
3221 
3222 	bge_stats_update(sc);
3223 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
3224 
3225 	if (sc->bge_flags & BGE_PHY_FIBER_TBI) {
3226 		if (CSR_READ_4(sc, BGE_MAC_STS) &
3227 		    BGE_MACSTAT_TBI_PCS_SYNCHED) {
3228 			CSR_WRITE_4(sc, BGE_MAC_STS, 0xFFFFFFFF);
3229 		}
3230 	} else
3231 		mii_tick(mii);
3232 
3233 	splx(s);
3234 }
3235 
3236 static void
3237 bge_stats_update(struct bge_softc *sc)
3238 {
3239 	struct ifnet *ifp = &sc->ethercom.ec_if;
3240 	bus_size_t stats = BGE_MEMWIN_START + BGE_STATS_BLOCK;
3241 	bus_size_t rstats = BGE_RX_STATS;
3242 
3243 #define READ_RSTAT(sc, stats, stat) \
3244 	  CSR_READ_4(sc, stats + offsetof(struct bge_mac_stats_regs, stat))
3245 
3246 	if (BGE_IS_5705_OR_BEYOND(sc)) {
3247 		ifp->if_collisions +=
3248 		    READ_RSTAT(sc, rstats, dot3StatsSingleCollisionFrames) +
3249 		    READ_RSTAT(sc, rstats, dot3StatsMultipleCollisionFrames) +
3250 		    READ_RSTAT(sc, rstats, dot3StatsExcessiveCollisions) +
3251 		    READ_RSTAT(sc, rstats, dot3StatsLateCollisions);
3252 
3253 		BGE_EVCNT_ADD(sc->bge_ev_tx_xoff,
3254 			      READ_RSTAT(sc, rstats, outXoffSent));
3255 		BGE_EVCNT_ADD(sc->bge_ev_tx_xon,
3256 			      READ_RSTAT(sc, rstats, outXonSent));
3257 		BGE_EVCNT_ADD(sc->bge_ev_rx_xoff,
3258 			      READ_RSTAT(sc, rstats, xoffPauseFramesReceived));
3259 		BGE_EVCNT_ADD(sc->bge_ev_rx_xon,
3260 			      READ_RSTAT(sc, rstats, xonPauseFramesReceived));
3261 		BGE_EVCNT_ADD(sc->bge_ev_rx_macctl,
3262 			      READ_RSTAT(sc, rstats, macControlFramesReceived));
3263 		BGE_EVCNT_ADD(sc->bge_ev_xoffentered,
3264 			      READ_RSTAT(sc, rstats, xoffStateEntered));
3265 		return;
3266 	}
3267 
3268 #undef READ_RSTAT
3269 #define READ_STAT(sc, stats, stat) \
3270 	  CSR_READ_4(sc, stats + offsetof(struct bge_stats, stat))
3271 
3272 	ifp->if_collisions +=
3273 	  (READ_STAT(sc, stats, dot3StatsSingleCollisionFrames.bge_addr_lo) +
3274 	   READ_STAT(sc, stats, dot3StatsMultipleCollisionFrames.bge_addr_lo) +
3275 	   READ_STAT(sc, stats, dot3StatsExcessiveCollisions.bge_addr_lo) +
3276 	   READ_STAT(sc, stats, dot3StatsLateCollisions.bge_addr_lo)) -
3277 	  ifp->if_collisions;
3278 
3279 	BGE_EVCNT_UPD(sc->bge_ev_tx_xoff,
3280 		      READ_STAT(sc, stats, outXoffSent.bge_addr_lo));
3281 	BGE_EVCNT_UPD(sc->bge_ev_tx_xon,
3282 		      READ_STAT(sc, stats, outXonSent.bge_addr_lo));
3283 	BGE_EVCNT_UPD(sc->bge_ev_rx_xoff,
3284 		      READ_STAT(sc, stats,
3285 		      		xoffPauseFramesReceived.bge_addr_lo));
3286 	BGE_EVCNT_UPD(sc->bge_ev_rx_xon,
3287 		      READ_STAT(sc, stats, xonPauseFramesReceived.bge_addr_lo));
3288 	BGE_EVCNT_UPD(sc->bge_ev_rx_macctl,
3289 		      READ_STAT(sc, stats,
3290 		      		macControlFramesReceived.bge_addr_lo));
3291 	BGE_EVCNT_UPD(sc->bge_ev_xoffentered,
3292 		      READ_STAT(sc, stats, xoffStateEntered.bge_addr_lo));
3293 
3294 #undef READ_STAT
3295 
3296 #ifdef notdef
3297 	ifp->if_collisions +=
3298 	   (sc->bge_rdata->bge_info.bge_stats.dot3StatsSingleCollisionFrames +
3299 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsMultipleCollisionFrames +
3300 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsExcessiveCollisions +
3301 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsLateCollisions) -
3302 	   ifp->if_collisions;
3303 #endif
3304 }
3305 
3306 /*
3307  * Pad outbound frame to ETHER_MIN_NOPAD for an unusual reason.
3308  * The bge hardware will pad out Tx runts to ETHER_MIN_NOPAD,
3309  * but when such padded frames employ the  bge IP/TCP checksum offload,
3310  * the hardware checksum assist gives incorrect results (possibly
3311  * from incorporating its own padding into the UDP/TCP checksum; who knows).
3312  * If we pad such runts with zeros, the onboard checksum comes out correct.
3313  */
3314 static inline int
3315 bge_cksum_pad(struct mbuf *pkt)
3316 {
3317 	struct mbuf *last = NULL;
3318 	int padlen;
3319 
3320 	padlen = ETHER_MIN_NOPAD - pkt->m_pkthdr.len;
3321 
3322 	/* if there's only the packet-header and we can pad there, use it. */
3323 	if (pkt->m_pkthdr.len == pkt->m_len &&
3324 	    M_TRAILINGSPACE(pkt) >= padlen) {
3325 		last = pkt;
3326 	} else {
3327 		/*
3328 		 * Walk packet chain to find last mbuf. We will either
3329 		 * pad there, or append a new mbuf and pad it
3330 		 * (thus perhaps avoiding the bcm5700 dma-min bug).
3331 		 */
3332 		for (last = pkt; last->m_next != NULL; last = last->m_next) {
3333 	      	       continue; /* do nothing */
3334 		}
3335 
3336 		/* `last' now points to last in chain. */
3337 		if (M_TRAILINGSPACE(last) < padlen) {
3338 			/* Allocate new empty mbuf, pad it. Compact later. */
3339 			struct mbuf *n;
3340 			MGET(n, M_DONTWAIT, MT_DATA);
3341 			if (n == NULL)
3342 				return ENOBUFS;
3343 			n->m_len = 0;
3344 			last->m_next = n;
3345 			last = n;
3346 		}
3347 	}
3348 
3349 	KDASSERT(!M_READONLY(last));
3350 	KDASSERT(M_TRAILINGSPACE(last) >= padlen);
3351 
3352 	/* Now zero the pad area, to avoid the bge cksum-assist bug */
3353 	memset(mtod(last, char *) + last->m_len, 0, padlen);
3354 	last->m_len += padlen;
3355 	pkt->m_pkthdr.len += padlen;
3356 	return 0;
3357 }
3358 
3359 /*
3360  * Compact outbound packets to avoid bug with DMA segments less than 8 bytes.
3361  */
3362 static inline int
3363 bge_compact_dma_runt(struct mbuf *pkt)
3364 {
3365 	struct mbuf	*m, *prev;
3366 	int 		totlen, prevlen;
3367 
3368 	prev = NULL;
3369 	totlen = 0;
3370 	prevlen = -1;
3371 
3372 	for (m = pkt; m != NULL; prev = m,m = m->m_next) {
3373 		int mlen = m->m_len;
3374 		int shortfall = 8 - mlen ;
3375 
3376 		totlen += mlen;
3377 		if (mlen == 0) {
3378 			continue;
3379 		}
3380 		if (mlen >= 8)
3381 			continue;
3382 
3383 		/* If we get here, mbuf data is too small for DMA engine.
3384 		 * Try to fix by shuffling data to prev or next in chain.
3385 		 * If that fails, do a compacting deep-copy of the whole chain.
3386 		 */
3387 
3388 		/* Internal frag. If fits in prev, copy it there. */
3389 		if (prev && M_TRAILINGSPACE(prev) >= m->m_len) {
3390 		  	memcpy(prev->m_data + prev->m_len, m->m_data, mlen);
3391 			prev->m_len += mlen;
3392 			m->m_len = 0;
3393 			/* XXX stitch chain */
3394 			prev->m_next = m_free(m);
3395 			m = prev;
3396 			continue;
3397 		}
3398 		else if (m->m_next != NULL &&
3399 			     M_TRAILINGSPACE(m) >= shortfall &&
3400 			     m->m_next->m_len >= (8 + shortfall)) {
3401 		    /* m is writable and have enough data in next, pull up. */
3402 
3403 		  	memcpy(m->m_data + m->m_len, m->m_next->m_data,
3404 			    shortfall);
3405 			m->m_len += shortfall;
3406 			m->m_next->m_len -= shortfall;
3407 			m->m_next->m_data += shortfall;
3408 		}
3409 		else if (m->m_next == NULL || 1) {
3410 		  	/* Got a runt at the very end of the packet.
3411 			 * borrow data from the tail of the preceding mbuf and
3412 			 * update its length in-place. (The original data is still
3413 			 * valid, so we can do this even if prev is not writable.)
3414 			 */
3415 
3416 			/* if we'd make prev a runt, just move all of its data. */
3417 			KASSERT(prev != NULL /*, ("runt but null PREV")*/);
3418 			KASSERT(prev->m_len >= 8 /*, ("runt prev")*/);
3419 
3420 			if ((prev->m_len - shortfall) < 8)
3421 				shortfall = prev->m_len;
3422 
3423 #ifdef notyet	/* just do the safe slow thing for now */
3424 			if (!M_READONLY(m)) {
3425 				if (M_LEADINGSPACE(m) < shorfall) {
3426 					void *m_dat;
3427 					m_dat = (m->m_flags & M_PKTHDR) ?
3428 					  m->m_pktdat : m->dat;
3429 					memmove(m_dat, mtod(m, void*), m->m_len);
3430 					m->m_data = m_dat;
3431 				    }
3432 			} else
3433 #endif	/* just do the safe slow thing */
3434 			{
3435 				struct mbuf * n = NULL;
3436 				int newprevlen = prev->m_len - shortfall;
3437 
3438 				MGET(n, M_NOWAIT, MT_DATA);
3439 				if (n == NULL)
3440 				   return ENOBUFS;
3441 				KASSERT(m->m_len + shortfall < MLEN
3442 					/*,
3443 					  ("runt %d +prev %d too big\n", m->m_len, shortfall)*/);
3444 
3445 				/* first copy the data we're stealing from prev */
3446 				memcpy(n->m_data, prev->m_data + newprevlen,
3447 				    shortfall);
3448 
3449 				/* update prev->m_len accordingly */
3450 				prev->m_len -= shortfall;
3451 
3452 				/* copy data from runt m */
3453 				memcpy(n->m_data + shortfall, m->m_data,
3454 				    m->m_len);
3455 
3456 				/* n holds what we stole from prev, plus m */
3457 				n->m_len = shortfall + m->m_len;
3458 
3459 				/* stitch n into chain and free m */
3460 				n->m_next = m->m_next;
3461 				prev->m_next = n;
3462 				/* KASSERT(m->m_next == NULL); */
3463 				m->m_next = NULL;
3464 				m_free(m);
3465 				m = n;	/* for continuing loop */
3466 			}
3467 		}
3468 		prevlen = m->m_len;
3469 	}
3470 	return 0;
3471 }
3472 
3473 /*
3474  * Encapsulate an mbuf chain in the tx ring  by coupling the mbuf data
3475  * pointers to descriptors.
3476  */
3477 static int
3478 bge_encap(struct bge_softc *sc, struct mbuf *m_head, u_int32_t *txidx)
3479 {
3480 	struct bge_tx_bd	*f = NULL;
3481 	u_int32_t		frag, cur;
3482 	u_int16_t		csum_flags = 0;
3483 	u_int16_t		txbd_tso_flags = 0;
3484 	struct txdmamap_pool_entry *dma;
3485 	bus_dmamap_t dmamap;
3486 	int			i = 0;
3487 	struct m_tag		*mtag;
3488 	int			use_tso, maxsegsize, error;
3489 
3490 	cur = frag = *txidx;
3491 
3492 	if (m_head->m_pkthdr.csum_flags) {
3493 		if (m_head->m_pkthdr.csum_flags & M_CSUM_IPv4)
3494 			csum_flags |= BGE_TXBDFLAG_IP_CSUM;
3495 		if (m_head->m_pkthdr.csum_flags & (M_CSUM_TCPv4|M_CSUM_UDPv4))
3496 			csum_flags |= BGE_TXBDFLAG_TCP_UDP_CSUM;
3497 	}
3498 
3499 	/*
3500 	 * If we were asked to do an outboard checksum, and the NIC
3501 	 * has the bug where it sometimes adds in the Ethernet padding,
3502 	 * explicitly pad with zeros so the cksum will be correct either way.
3503 	 * (For now, do this for all chip versions, until newer
3504 	 * are confirmed to not require the workaround.)
3505 	 */
3506 	if ((csum_flags & BGE_TXBDFLAG_TCP_UDP_CSUM) == 0 ||
3507 #ifdef notyet
3508 	    (sc->bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||
3509 #endif
3510 	    m_head->m_pkthdr.len >= ETHER_MIN_NOPAD)
3511 		goto check_dma_bug;
3512 
3513 	if (bge_cksum_pad(m_head) != 0) {
3514 	    return ENOBUFS;
3515 	}
3516 
3517 check_dma_bug:
3518 	if (!(BGE_CHIPREV(sc->bge_chipid) == BGE_CHIPREV_5700_BX))
3519 		goto doit;
3520 
3521 	/*
3522 	 * bcm5700 Revision B silicon cannot handle DMA descriptors with
3523 	 * less than eight bytes.  If we encounter a teeny mbuf
3524 	 * at the end of a chain, we can pad.  Otherwise, copy.
3525 	 */
3526 	if (bge_compact_dma_runt(m_head) != 0)
3527 		return ENOBUFS;
3528 
3529 doit:
3530 	dma = SLIST_FIRST(&sc->txdma_list);
3531 	if (dma == NULL)
3532 		return ENOBUFS;
3533 	dmamap = dma->dmamap;
3534 
3535 	/*
3536 	 * Set up any necessary TSO state before we start packing...
3537 	 */
3538 	use_tso = (m_head->m_pkthdr.csum_flags & M_CSUM_TSOv4) != 0;
3539 	if (!use_tso) {
3540 		maxsegsize = 0;
3541 	} else {	/* TSO setup */
3542 		unsigned  mss;
3543 		struct ether_header *eh;
3544 		unsigned ip_tcp_hlen, iptcp_opt_words, tcp_seg_flags, offset;
3545 		struct mbuf * m0 = m_head;
3546 		struct ip *ip;
3547 		struct tcphdr *th;
3548 		int iphl, hlen;
3549 
3550 		/*
3551 		 * XXX It would be nice if the mbuf pkthdr had offset
3552 		 * fields for the protocol headers.
3553 		 */
3554 
3555 		eh = mtod(m0, struct ether_header *);
3556 		switch (htons(eh->ether_type)) {
3557 		case ETHERTYPE_IP:
3558 			offset = ETHER_HDR_LEN;
3559 			break;
3560 
3561 		case ETHERTYPE_VLAN:
3562 			offset = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3563 			break;
3564 
3565 		default:
3566 			/*
3567 			 * Don't support this protocol or encapsulation.
3568 			 */
3569 			return (ENOBUFS);
3570 		}
3571 
3572 		/*
3573 		 * TCP/IP headers are in the first mbuf; we can do
3574 		 * this the easy way.
3575 		 */
3576 		iphl = M_CSUM_DATA_IPv4_IPHL(m0->m_pkthdr.csum_data);
3577 		hlen = iphl + offset;
3578 		if (__predict_false(m0->m_len <
3579 				    (hlen + sizeof(struct tcphdr)))) {
3580 
3581 			aprint_debug_dev(sc->bge_dev,
3582 			    "TSO: hard case m0->m_len == %d < ip/tcp hlen %zd,"
3583 			    "not handled yet\n",
3584 			     m0->m_len, hlen+ sizeof(struct tcphdr));
3585 #ifdef NOTYET
3586 			/*
3587 			 * XXX jonathan@NetBSD.org: untested.
3588 			 * how to force  this branch to be taken?
3589 			 */
3590 			BGE_EVCNT_INCR(&sc->sc_ev_txtsopain);
3591 
3592 			m_copydata(m0, offset, sizeof(ip), &ip);
3593 			m_copydata(m0, hlen, sizeof(th), &th);
3594 
3595 			ip.ip_len = 0;
3596 
3597 			m_copyback(m0, hlen + offsetof(struct ip, ip_len),
3598 			    sizeof(ip.ip_len), &ip.ip_len);
3599 
3600 			th.th_sum = in_cksum_phdr(ip.ip_src.s_addr,
3601 			    ip.ip_dst.s_addr, htons(IPPROTO_TCP));
3602 
3603 			m_copyback(m0, hlen + offsetof(struct tcphdr, th_sum),
3604 			    sizeof(th.th_sum), &th.th_sum);
3605 
3606 			hlen += th.th_off << 2;
3607 			iptcp_opt_words	= hlen;
3608 #else
3609 			/*
3610 			 * if_wm "hard" case not yet supported, can we not
3611 			 * mandate it out of existence?
3612 			 */
3613 			(void) ip; (void)th; (void) ip_tcp_hlen;
3614 
3615 			return ENOBUFS;
3616 #endif
3617 		} else {
3618 			ip = (struct ip *) (mtod(m0, char *) + offset);
3619 			th = (struct tcphdr *) (mtod(m0, char *) + hlen);
3620 			ip_tcp_hlen = iphl +  (th->th_off << 2);
3621 
3622 			/* Total IP/TCP options, in 32-bit words */
3623 			iptcp_opt_words = (ip_tcp_hlen
3624 					   - sizeof(struct tcphdr)
3625 					   - sizeof(struct ip)) >> 2;
3626 		}
3627 		if (BGE_IS_5750_OR_BEYOND(sc)) {
3628 			th->th_sum = 0;
3629 			csum_flags &= ~(BGE_TXBDFLAG_TCP_UDP_CSUM);
3630 		} else {
3631 			/*
3632 			 * XXX jonathan@NetBSD.org: 5705 untested.
3633 			 * Requires TSO firmware patch for 5701/5703/5704.
3634 			 */
3635 			th->th_sum = in_cksum_phdr(ip->ip_src.s_addr,
3636 			    ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3637 		}
3638 
3639 		mss = m_head->m_pkthdr.segsz;
3640 		txbd_tso_flags |=
3641 		    BGE_TXBDFLAG_CPU_PRE_DMA |
3642 		    BGE_TXBDFLAG_CPU_POST_DMA;
3643 
3644 		/*
3645 		 * Our NIC TSO-assist assumes TSO has standard, optionless
3646 		 * IPv4 and TCP headers, which total 40 bytes. By default,
3647 		 * the NIC copies 40 bytes of IP/TCP header from the
3648 		 * supplied header into the IP/TCP header portion of
3649 		 * each post-TSO-segment. If the supplied packet has IP or
3650 		 * TCP options, we need to tell the NIC to copy those extra
3651 		 * bytes into each  post-TSO header, in addition to the normal
3652 		 * 40-byte IP/TCP header (and to leave space accordingly).
3653 		 * Unfortunately, the driver encoding of option length
3654 		 * varies across different ASIC families.
3655 		 */
3656 		tcp_seg_flags = 0;
3657 		if (iptcp_opt_words) {
3658 			if ( BGE_IS_5705_OR_BEYOND(sc)) {
3659 				tcp_seg_flags =
3660 					iptcp_opt_words << 11;
3661 			} else {
3662 				txbd_tso_flags |=
3663 					iptcp_opt_words << 12;
3664 			}
3665 		}
3666 		maxsegsize = mss | tcp_seg_flags;
3667 		ip->ip_len = htons(mss + ip_tcp_hlen);
3668 
3669 	}	/* TSO setup */
3670 
3671 	/*
3672 	 * Start packing the mbufs in this chain into
3673 	 * the fragment pointers. Stop when we run out
3674 	 * of fragments or hit the end of the mbuf chain.
3675 	 */
3676 	error = bus_dmamap_load_mbuf(sc->bge_dmatag, dmamap, m_head,
3677 	    BUS_DMA_NOWAIT);
3678 	if (error) {
3679 		return (ENOBUFS);
3680 	}
3681 	/*
3682 	 * Sanity check: avoid coming within 16 descriptors
3683 	 * of the end of the ring.
3684 	 */
3685 	if (dmamap->dm_nsegs > (BGE_TX_RING_CNT - sc->bge_txcnt - 16)) {
3686 		BGE_TSO_PRINTF(("%s: "
3687 		    " dmamap_load_mbuf too close to ring wrap\n",
3688 		    device_xname(sc->bge_dev)));
3689 		goto fail_unload;
3690 	}
3691 
3692 	mtag = sc->ethercom.ec_nvlans ?
3693 	    m_tag_find(m_head, PACKET_TAG_VLAN, NULL) : NULL;
3694 
3695 
3696 	/* Iterate over dmap-map fragments. */
3697 	for (i = 0; i < dmamap->dm_nsegs; i++) {
3698 		f = &sc->bge_rdata->bge_tx_ring[frag];
3699 		if (sc->bge_cdata.bge_tx_chain[frag] != NULL)
3700 			break;
3701 
3702 		bge_set_hostaddr(&f->bge_addr, dmamap->dm_segs[i].ds_addr);
3703 		f->bge_len = dmamap->dm_segs[i].ds_len;
3704 
3705 		/*
3706 		 * For 5751 and follow-ons, for TSO we must turn
3707 		 * off checksum-assist flag in the tx-descr, and
3708 		 * supply the ASIC-revision-specific encoding
3709 		 * of TSO flags and segsize.
3710 		 */
3711 		if (use_tso) {
3712 			if (BGE_IS_5750_OR_BEYOND(sc) || i == 0) {
3713 				f->bge_rsvd = maxsegsize;
3714 				f->bge_flags = csum_flags | txbd_tso_flags;
3715 			} else {
3716 				f->bge_rsvd = 0;
3717 				f->bge_flags =
3718 				  (csum_flags | txbd_tso_flags) & 0x0fff;
3719 			}
3720 		} else {
3721 			f->bge_rsvd = 0;
3722 			f->bge_flags = csum_flags;
3723 		}
3724 
3725 		if (mtag != NULL) {
3726 			f->bge_flags |= BGE_TXBDFLAG_VLAN_TAG;
3727 			f->bge_vlan_tag = VLAN_TAG_VALUE(mtag);
3728 		} else {
3729 			f->bge_vlan_tag = 0;
3730 		}
3731 		cur = frag;
3732 		BGE_INC(frag, BGE_TX_RING_CNT);
3733 	}
3734 
3735 	if (i < dmamap->dm_nsegs) {
3736 		BGE_TSO_PRINTF(("%s: reached %d < dm_nsegs %d\n",
3737 		    device_xname(sc->bge_dev), i, dmamap->dm_nsegs));
3738 		goto fail_unload;
3739 	}
3740 
3741 	bus_dmamap_sync(sc->bge_dmatag, dmamap, 0, dmamap->dm_mapsize,
3742 	    BUS_DMASYNC_PREWRITE);
3743 
3744 	if (frag == sc->bge_tx_saved_considx) {
3745 		BGE_TSO_PRINTF(("%s: frag %d = wrapped id %d?\n",
3746 		    device_xname(sc->bge_dev), frag, sc->bge_tx_saved_considx));
3747 
3748 		goto fail_unload;
3749 	}
3750 
3751 	sc->bge_rdata->bge_tx_ring[cur].bge_flags |= BGE_TXBDFLAG_END;
3752 	sc->bge_cdata.bge_tx_chain[cur] = m_head;
3753 	SLIST_REMOVE_HEAD(&sc->txdma_list, link);
3754 	sc->txdma[cur] = dma;
3755 	sc->bge_txcnt += dmamap->dm_nsegs;
3756 
3757 	*txidx = frag;
3758 
3759 	return (0);
3760 
3761 fail_unload:
3762 	bus_dmamap_unload(sc->bge_dmatag, dmamap);
3763 
3764 	return ENOBUFS;
3765 }
3766 
3767 /*
3768  * Main transmit routine. To avoid having to do mbuf copies, we put pointers
3769  * to the mbuf data regions directly in the transmit descriptors.
3770  */
3771 static void
3772 bge_start(struct ifnet *ifp)
3773 {
3774 	struct bge_softc *sc;
3775 	struct mbuf *m_head = NULL;
3776 	u_int32_t prodidx;
3777 	int pkts = 0;
3778 
3779 	sc = ifp->if_softc;
3780 
3781 	if ((ifp->if_flags & (IFF_RUNNING|IFF_OACTIVE)) != IFF_RUNNING)
3782 		return;
3783 
3784 	prodidx = sc->bge_tx_prodidx;
3785 
3786 	while(sc->bge_cdata.bge_tx_chain[prodidx] == NULL) {
3787 		IFQ_POLL(&ifp->if_snd, m_head);
3788 		if (m_head == NULL)
3789 			break;
3790 
3791 #if 0
3792 		/*
3793 		 * XXX
3794 		 * safety overkill.  If this is a fragmented packet chain
3795 		 * with delayed TCP/UDP checksums, then only encapsulate
3796 		 * it if we have enough descriptors to handle the entire
3797 		 * chain at once.
3798 		 * (paranoia -- may not actually be needed)
3799 		 */
3800 		if (m_head->m_flags & M_FIRSTFRAG &&
3801 		    m_head->m_pkthdr.csum_flags & (CSUM_DELAY_DATA)) {
3802 			if ((BGE_TX_RING_CNT - sc->bge_txcnt) <
3803 			    M_CSUM_DATA_IPv4_OFFSET(m_head->m_pkthdr.csum_data) + 16) {
3804 				ifp->if_flags |= IFF_OACTIVE;
3805 				break;
3806 			}
3807 		}
3808 #endif
3809 
3810 		/*
3811 		 * Pack the data into the transmit ring. If we
3812 		 * don't have room, set the OACTIVE flag and wait
3813 		 * for the NIC to drain the ring.
3814 		 */
3815 		if (bge_encap(sc, m_head, &prodidx)) {
3816 			ifp->if_flags |= IFF_OACTIVE;
3817 			break;
3818 		}
3819 
3820 		/* now we are committed to transmit the packet */
3821 		IFQ_DEQUEUE(&ifp->if_snd, m_head);
3822 		pkts++;
3823 
3824 #if NBPFILTER > 0
3825 		/*
3826 		 * If there's a BPF listener, bounce a copy of this frame
3827 		 * to him.
3828 		 */
3829 		if (ifp->if_bpf)
3830 			bpf_mtap(ifp->if_bpf, m_head);
3831 #endif
3832 	}
3833 	if (pkts == 0)
3834 		return;
3835 
3836 	/* Transmit */
3837 	bge_writembx(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
3838 	/* 5700 b2 errata */
3839 	if (BGE_CHIPREV(sc->bge_chipid) == BGE_CHIPREV_5700_BX)
3840 		bge_writembx(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
3841 
3842 	sc->bge_tx_prodidx = prodidx;
3843 
3844 	/*
3845 	 * Set a timeout in case the chip goes out to lunch.
3846 	 */
3847 	ifp->if_timer = 5;
3848 }
3849 
3850 static int
3851 bge_init(struct ifnet *ifp)
3852 {
3853 	struct bge_softc *sc = ifp->if_softc;
3854 	const u_int16_t *m;
3855 	int s, error = 0;
3856 
3857 	s = splnet();
3858 
3859 	ifp = &sc->ethercom.ec_if;
3860 
3861 	/* Cancel pending I/O and flush buffers. */
3862 	bge_stop(ifp, 0);
3863 	bge_reset(sc);
3864 	bge_chipinit(sc);
3865 
3866 	/*
3867 	 * Init the various state machines, ring
3868 	 * control blocks and firmware.
3869 	 */
3870 	error = bge_blockinit(sc);
3871 	if (error != 0) {
3872 		aprint_error_dev(sc->bge_dev, "initialization error %d\n",
3873 		    error);
3874 		splx(s);
3875 		return error;
3876 	}
3877 
3878 	ifp = &sc->ethercom.ec_if;
3879 
3880 	/* Specify MTU. */
3881 	CSR_WRITE_4(sc, BGE_RX_MTU, ifp->if_mtu +
3882 	    ETHER_HDR_LEN + ETHER_CRC_LEN + ETHER_VLAN_ENCAP_LEN);
3883 
3884 	/* Load our MAC address. */
3885 	m = (const u_int16_t *)&(CLLADDR(ifp->if_sadl)[0]);
3886 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_LO, htons(m[0]));
3887 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_HI, (htons(m[1]) << 16) | htons(m[2]));
3888 
3889 	/* Enable or disable promiscuous mode as needed. */
3890 	if (ifp->if_flags & IFF_PROMISC) {
3891 		BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
3892 	} else {
3893 		BGE_CLRBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
3894 	}
3895 
3896 	/* Program multicast filter. */
3897 	bge_setmulti(sc);
3898 
3899 	/* Init RX ring. */
3900 	bge_init_rx_ring_std(sc);
3901 
3902 	/* Init jumbo RX ring. */
3903 	if (ifp->if_mtu > (ETHERMTU + ETHER_HDR_LEN + ETHER_CRC_LEN))
3904 		bge_init_rx_ring_jumbo(sc);
3905 
3906 	/* Init our RX return ring index */
3907 	sc->bge_rx_saved_considx = 0;
3908 
3909 	/* Init TX ring. */
3910 	bge_init_tx_ring(sc);
3911 
3912 	/* Turn on transmitter */
3913 	BGE_SETBIT(sc, BGE_TX_MODE, BGE_TXMODE_ENABLE);
3914 
3915 	/* Turn on receiver */
3916 	BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
3917 
3918 	CSR_WRITE_4(sc, BGE_MAX_RX_FRAME_LOWAT, 2);
3919 
3920 	/* Tell firmware we're alive. */
3921 	BGE_SETBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
3922 
3923 	/* Enable host interrupts. */
3924 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_CLEAR_INTA);
3925 	BGE_CLRBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
3926 	bge_writembx(sc, BGE_MBX_IRQ0_LO, 0);
3927 
3928 	if ((error = bge_ifmedia_upd(ifp)) != 0)
3929 		goto out;
3930 
3931 	ifp->if_flags |= IFF_RUNNING;
3932 	ifp->if_flags &= ~IFF_OACTIVE;
3933 
3934 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
3935 
3936 out:
3937 	splx(s);
3938 
3939 	return error;
3940 }
3941 
3942 /*
3943  * Set media options.
3944  */
3945 static int
3946 bge_ifmedia_upd(struct ifnet *ifp)
3947 {
3948 	struct bge_softc *sc = ifp->if_softc;
3949 	struct mii_data *mii = &sc->bge_mii;
3950 	struct ifmedia *ifm = &sc->bge_ifmedia;
3951 	int rc;
3952 
3953 	/* If this is a 1000baseX NIC, enable the TBI port. */
3954 	if (sc->bge_flags & BGE_PHY_FIBER_TBI) {
3955 		if (IFM_TYPE(ifm->ifm_media) != IFM_ETHER)
3956 			return (EINVAL);
3957 		switch(IFM_SUBTYPE(ifm->ifm_media)) {
3958 		case IFM_AUTO:
3959 			break;
3960 		case IFM_1000_SX:
3961 			if ((ifm->ifm_media & IFM_GMASK) == IFM_FDX) {
3962 				BGE_CLRBIT(sc, BGE_MAC_MODE,
3963 				    BGE_MACMODE_HALF_DUPLEX);
3964 			} else {
3965 				BGE_SETBIT(sc, BGE_MAC_MODE,
3966 				    BGE_MACMODE_HALF_DUPLEX);
3967 			}
3968 			break;
3969 		default:
3970 			return (EINVAL);
3971 		}
3972 		/* XXX 802.3x flow control for 1000BASE-SX */
3973 		return (0);
3974 	}
3975 
3976 	sc->bge_link = 0;
3977 	if ((rc = mii_mediachg(mii)) == ENXIO)
3978 		return 0;
3979 	return rc;
3980 }
3981 
3982 /*
3983  * Report current media status.
3984  */
3985 static void
3986 bge_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
3987 {
3988 	struct bge_softc *sc = ifp->if_softc;
3989 	struct mii_data *mii = &sc->bge_mii;
3990 
3991 	if (sc->bge_flags & BGE_PHY_FIBER_TBI) {
3992 		ifmr->ifm_status = IFM_AVALID;
3993 		ifmr->ifm_active = IFM_ETHER;
3994 		if (CSR_READ_4(sc, BGE_MAC_STS) &
3995 		    BGE_MACSTAT_TBI_PCS_SYNCHED)
3996 			ifmr->ifm_status |= IFM_ACTIVE;
3997 		ifmr->ifm_active |= IFM_1000_SX;
3998 		if (CSR_READ_4(sc, BGE_MAC_MODE) & BGE_MACMODE_HALF_DUPLEX)
3999 			ifmr->ifm_active |= IFM_HDX;
4000 		else
4001 			ifmr->ifm_active |= IFM_FDX;
4002 		return;
4003 	}
4004 
4005 	mii_pollstat(mii);
4006 	ifmr->ifm_status = mii->mii_media_status;
4007 	ifmr->ifm_active = (mii->mii_media_active & ~IFM_ETH_FMASK) |
4008 	    sc->bge_flowflags;
4009 }
4010 
4011 static int
4012 bge_ioctl(struct ifnet *ifp, u_long command, void *data)
4013 {
4014 	struct bge_softc *sc = ifp->if_softc;
4015 	struct ifreq *ifr = (struct ifreq *) data;
4016 	int s, error = 0;
4017 	struct mii_data *mii;
4018 
4019 	s = splnet();
4020 
4021 	switch(command) {
4022 	case SIOCSIFFLAGS:
4023 		if ((error = ifioctl_common(ifp, command, data)) != 0)
4024 			break;
4025 		if (ifp->if_flags & IFF_UP) {
4026 			/*
4027 			 * If only the state of the PROMISC flag changed,
4028 			 * then just use the 'set promisc mode' command
4029 			 * instead of reinitializing the entire NIC. Doing
4030 			 * a full re-init means reloading the firmware and
4031 			 * waiting for it to start up, which may take a
4032 			 * second or two.
4033 			 */
4034 			if (ifp->if_flags & IFF_RUNNING &&
4035 			    ifp->if_flags & IFF_PROMISC &&
4036 			    !(sc->bge_if_flags & IFF_PROMISC)) {
4037 				BGE_SETBIT(sc, BGE_RX_MODE,
4038 				    BGE_RXMODE_RX_PROMISC);
4039 			} else if (ifp->if_flags & IFF_RUNNING &&
4040 			    !(ifp->if_flags & IFF_PROMISC) &&
4041 			    sc->bge_if_flags & IFF_PROMISC) {
4042 				BGE_CLRBIT(sc, BGE_RX_MODE,
4043 				    BGE_RXMODE_RX_PROMISC);
4044 			} else if (!(sc->bge_if_flags & IFF_UP))
4045 				bge_init(ifp);
4046 		} else {
4047 			if (ifp->if_flags & IFF_RUNNING)
4048 				bge_stop(ifp, 1);
4049 		}
4050 		sc->bge_if_flags = ifp->if_flags;
4051 		error = 0;
4052 		break;
4053 	case SIOCSIFMEDIA:
4054 		/* XXX Flow control is not supported for 1000BASE-SX */
4055 		if (sc->bge_flags & BGE_PHY_FIBER_TBI) {
4056 			ifr->ifr_media &= ~IFM_ETH_FMASK;
4057 			sc->bge_flowflags = 0;
4058 		}
4059 
4060 		/* Flow control requires full-duplex mode. */
4061 		if (IFM_SUBTYPE(ifr->ifr_media) == IFM_AUTO ||
4062 		    (ifr->ifr_media & IFM_FDX) == 0) {
4063 		    	ifr->ifr_media &= ~IFM_ETH_FMASK;
4064 		}
4065 		if (IFM_SUBTYPE(ifr->ifr_media) != IFM_AUTO) {
4066 			if ((ifr->ifr_media & IFM_ETH_FMASK) == IFM_FLOW) {
4067 				/* We can do both TXPAUSE and RXPAUSE. */
4068 				ifr->ifr_media |=
4069 				    IFM_ETH_TXPAUSE | IFM_ETH_RXPAUSE;
4070 			}
4071 			sc->bge_flowflags = ifr->ifr_media & IFM_ETH_FMASK;
4072 		}
4073 		/* FALLTHROUGH */
4074 	case SIOCGIFMEDIA:
4075 		if (sc->bge_flags & BGE_PHY_FIBER_TBI) {
4076 			error = ifmedia_ioctl(ifp, ifr, &sc->bge_ifmedia,
4077 			    command);
4078 		} else {
4079 			mii = &sc->bge_mii;
4080 			error = ifmedia_ioctl(ifp, ifr, &mii->mii_media,
4081 			    command);
4082 		}
4083 		break;
4084 	default:
4085 		if ((error = ether_ioctl(ifp, command, data)) != ENETRESET)
4086 			break;
4087 
4088 		error = 0;
4089 
4090 		if (command != SIOCADDMULTI && command != SIOCDELMULTI)
4091 			;
4092 		else if (ifp->if_flags & IFF_RUNNING)
4093 			bge_setmulti(sc);
4094 		break;
4095 	}
4096 
4097 	splx(s);
4098 
4099 	return (error);
4100 }
4101 
4102 static void
4103 bge_watchdog(struct ifnet *ifp)
4104 {
4105 	struct bge_softc *sc;
4106 
4107 	sc = ifp->if_softc;
4108 
4109 	aprint_error_dev(sc->bge_dev, "watchdog timeout -- resetting\n");
4110 
4111 	ifp->if_flags &= ~IFF_RUNNING;
4112 	bge_init(ifp);
4113 
4114 	ifp->if_oerrors++;
4115 }
4116 
4117 static void
4118 bge_stop_block(struct bge_softc *sc, bus_addr_t reg, uint32_t bit)
4119 {
4120 	int i;
4121 
4122 	BGE_CLRBIT(sc, reg, bit);
4123 
4124 	for (i = 0; i < BGE_TIMEOUT; i++) {
4125 		if ((CSR_READ_4(sc, reg) & bit) == 0)
4126 			return;
4127 		delay(100);
4128 		if (sc->bge_flags & BGE_PCIE)
4129 		  DELAY(1000);
4130 	}
4131 
4132 	aprint_error_dev(sc->bge_dev,
4133 	    "block failed to stop: reg 0x%lx, bit 0x%08x\n", (u_long)reg, bit);
4134 }
4135 
4136 /*
4137  * Stop the adapter and free any mbufs allocated to the
4138  * RX and TX lists.
4139  */
4140 static void
4141 bge_stop(struct ifnet *ifp, int disable)
4142 {
4143 	struct bge_softc *sc = ifp->if_softc;
4144 
4145 	callout_stop(&sc->bge_timeout);
4146 
4147 	/*
4148 	 * Disable all of the receiver blocks
4149 	 */
4150 	bge_stop_block(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
4151 	bge_stop_block(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
4152 	bge_stop_block(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
4153 	if (!(BGE_IS_5705_OR_BEYOND(sc)))
4154 		bge_stop_block(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
4155 	bge_stop_block(sc, BGE_RDBDI_MODE, BGE_RBDIMODE_ENABLE);
4156 	bge_stop_block(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
4157 	bge_stop_block(sc, BGE_RBDC_MODE, BGE_RBDCMODE_ENABLE);
4158 
4159 	/*
4160 	 * Disable all of the transmit blocks
4161 	 */
4162 	bge_stop_block(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
4163 	bge_stop_block(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
4164 	bge_stop_block(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
4165 	bge_stop_block(sc, BGE_RDMA_MODE, BGE_RDMAMODE_ENABLE);
4166 	bge_stop_block(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
4167 	if (!(BGE_IS_5705_OR_BEYOND(sc)))
4168 		bge_stop_block(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
4169 	bge_stop_block(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
4170 
4171 	/*
4172 	 * Shut down all of the memory managers and related
4173 	 * state machines.
4174 	 */
4175 	bge_stop_block(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
4176 	bge_stop_block(sc, BGE_WDMA_MODE, BGE_WDMAMODE_ENABLE);
4177 	if (!(BGE_IS_5705_OR_BEYOND(sc)))
4178 		bge_stop_block(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
4179 
4180 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0xFFFFFFFF);
4181 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0);
4182 
4183 	if (!(BGE_IS_5705_OR_BEYOND(sc))) {
4184 		bge_stop_block(sc, BGE_BMAN_MODE, BGE_BMANMODE_ENABLE);
4185 		bge_stop_block(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE);
4186 	}
4187 
4188 	/* Disable host interrupts. */
4189 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
4190 	bge_writembx(sc, BGE_MBX_IRQ0_LO, 1);
4191 
4192 	/*
4193 	 * Tell firmware we're shutting down.
4194 	 */
4195 	BGE_CLRBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
4196 
4197 	/* Free the RX lists. */
4198 	bge_free_rx_ring_std(sc);
4199 
4200 	/* Free jumbo RX list. */
4201 	bge_free_rx_ring_jumbo(sc);
4202 
4203 	/* Free TX buffers. */
4204 	bge_free_tx_ring(sc);
4205 
4206 	/*
4207 	 * Isolate/power down the PHY.
4208 	 */
4209 	if (!(sc->bge_flags & BGE_PHY_FIBER_TBI))
4210 		mii_down(&sc->bge_mii);
4211 
4212 	sc->bge_link = 0;
4213 
4214 	sc->bge_tx_saved_considx = BGE_TXCONS_UNSET;
4215 
4216 	ifp->if_flags &= ~(IFF_RUNNING | IFF_OACTIVE);
4217 }
4218 
4219 static int
4220 sysctl_bge_verify(SYSCTLFN_ARGS)
4221 {
4222 	int error, t;
4223 	struct sysctlnode node;
4224 
4225 	node = *rnode;
4226 	t = *(int*)rnode->sysctl_data;
4227 	node.sysctl_data = &t;
4228 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
4229 	if (error || newp == NULL)
4230 		return (error);
4231 
4232 #if 0
4233 	DPRINTF2(("%s: t = %d, nodenum = %d, rnodenum = %d\n", __func__, t,
4234 	    node.sysctl_num, rnode->sysctl_num));
4235 #endif
4236 
4237 	if (node.sysctl_num == bge_rxthresh_nodenum) {
4238 		if (t < 0 || t >= NBGE_RX_THRESH)
4239 			return (EINVAL);
4240 		bge_update_all_threshes(t);
4241 	} else
4242 		return (EINVAL);
4243 
4244 	*(int*)rnode->sysctl_data = t;
4245 
4246 	return (0);
4247 }
4248 
4249 /*
4250  * Set up sysctl(3) MIB, hw.bge.*.
4251  *
4252  * TBD condition SYSCTL_PERMANENT on being an LKM or not
4253  */
4254 SYSCTL_SETUP(sysctl_bge, "sysctl bge subtree setup")
4255 {
4256 	int rc, bge_root_num;
4257 	const struct sysctlnode *node;
4258 
4259 	if ((rc = sysctl_createv(clog, 0, NULL, NULL,
4260 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "hw", NULL,
4261 	    NULL, 0, NULL, 0, CTL_HW, CTL_EOL)) != 0) {
4262 		goto err;
4263 	}
4264 
4265 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
4266 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "bge",
4267 	    SYSCTL_DESCR("BGE interface controls"),
4268 	    NULL, 0, NULL, 0, CTL_HW, CTL_CREATE, CTL_EOL)) != 0) {
4269 		goto err;
4270 	}
4271 
4272 	bge_root_num = node->sysctl_num;
4273 
4274 	/* BGE Rx interrupt mitigation level */
4275 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
4276 	    CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
4277 	    CTLTYPE_INT, "rx_lvl",
4278 	    SYSCTL_DESCR("BGE receive interrupt mitigation level"),
4279 	    sysctl_bge_verify, 0,
4280 	    &bge_rx_thresh_lvl,
4281 	    0, CTL_HW, bge_root_num, CTL_CREATE,
4282 	    CTL_EOL)) != 0) {
4283 		goto err;
4284 	}
4285 
4286 	bge_rxthresh_nodenum = node->sysctl_num;
4287 
4288 	return;
4289 
4290 err:
4291 	aprint_error("%s: sysctl_createv failed (rc = %d)\n", __func__, rc);
4292 }
4293 
4294 static int
4295 bge_get_eaddr_mem(struct bge_softc *sc, u_int8_t ether_addr[])
4296 {
4297 	u_int32_t mac_addr;
4298 
4299 	mac_addr = bge_readmem_ind(sc, 0x0c14);
4300 	if ((mac_addr >> 16) == 0x484b) {
4301 		ether_addr[0] = (uint8_t)(mac_addr >> 8);
4302 		ether_addr[1] = (uint8_t)mac_addr;
4303 		mac_addr = bge_readmem_ind(sc, 0x0c18);
4304 		ether_addr[2] = (uint8_t)(mac_addr >> 24);
4305 		ether_addr[3] = (uint8_t)(mac_addr >> 16);
4306 		ether_addr[4] = (uint8_t)(mac_addr >> 8);
4307 		ether_addr[5] = (uint8_t)mac_addr;
4308 		return (0);
4309 	}
4310 	return (1);
4311 }
4312 
4313 static int
4314 bge_get_eaddr_nvram(struct bge_softc *sc, u_int8_t ether_addr[])
4315 {
4316 	int mac_offset = BGE_EE_MAC_OFFSET;
4317 
4318 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5906) {
4319 		mac_offset = BGE_EE_MAC_OFFSET_5906;
4320 	}
4321 
4322 	return (bge_read_nvram(sc, ether_addr, mac_offset + 2,
4323 	    ETHER_ADDR_LEN));
4324 }
4325 
4326 static int
4327 bge_get_eaddr_eeprom(struct bge_softc *sc, u_int8_t ether_addr[])
4328 {
4329 
4330 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5906) {
4331 		return (1);
4332 	}
4333 
4334 	return (bge_read_eeprom(sc, ether_addr, BGE_EE_MAC_OFFSET + 2,
4335 	   ETHER_ADDR_LEN));
4336 }
4337 
4338 static int
4339 bge_get_eaddr(struct bge_softc *sc, u_int8_t eaddr[])
4340 {
4341 	static const bge_eaddr_fcn_t bge_eaddr_funcs[] = {
4342 		/* NOTE: Order is critical */
4343 		bge_get_eaddr_mem,
4344 		bge_get_eaddr_nvram,
4345 		bge_get_eaddr_eeprom,
4346 		NULL
4347 	};
4348 	const bge_eaddr_fcn_t *func;
4349 
4350 	for (func = bge_eaddr_funcs; *func != NULL; ++func) {
4351 		if ((*func)(sc, eaddr) == 0)
4352 			break;
4353 	}
4354 	return (*func == NULL ? ENXIO : 0);
4355 }
4356