xref: /netbsd-src/sys/dev/pci/if_bge.c (revision cac8e449158efc7261bebc8657cbb0125a2cfdde)
1 /*	$NetBSD: if_bge.c,v 1.150 2008/07/25 19:45:06 dsl 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 1.150 2008/07/25 19:45:06 dsl 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 static int	bge_probe(device_t, cfdata_t, void *);
190 static void	bge_attach(device_t, device_t, void *);
191 static void	bge_release_resources(struct bge_softc *);
192 static void	bge_txeof(struct bge_softc *);
193 static void	bge_rxeof(struct bge_softc *);
194 
195 static void	bge_tick(void *);
196 static void	bge_stats_update(struct bge_softc *);
197 static int	bge_encap(struct bge_softc *, struct mbuf *, u_int32_t *);
198 
199 static int	bge_intr(void *);
200 static void	bge_start(struct ifnet *);
201 static int	bge_ioctl(struct ifnet *, u_long, void *);
202 static int	bge_init(struct ifnet *);
203 static void	bge_stop(struct ifnet *, int);
204 static void	bge_watchdog(struct ifnet *);
205 static int	bge_ifmedia_upd(struct ifnet *);
206 static void	bge_ifmedia_sts(struct ifnet *, struct ifmediareq *);
207 
208 static void	bge_setmulti(struct bge_softc *);
209 
210 static void	bge_handle_events(struct bge_softc *);
211 static int	bge_alloc_jumbo_mem(struct bge_softc *);
212 #if 0 /* XXX */
213 static void	bge_free_jumbo_mem(struct bge_softc *);
214 #endif
215 static void	*bge_jalloc(struct bge_softc *);
216 static void	bge_jfree(struct mbuf *, void *, size_t, void *);
217 static int	bge_newbuf_std(struct bge_softc *, int, struct mbuf *,
218 			       bus_dmamap_t);
219 static int	bge_newbuf_jumbo(struct bge_softc *, int, struct mbuf *);
220 static int	bge_init_rx_ring_std(struct bge_softc *);
221 static void	bge_free_rx_ring_std(struct bge_softc *);
222 static int	bge_init_rx_ring_jumbo(struct bge_softc *);
223 static void	bge_free_rx_ring_jumbo(struct bge_softc *);
224 static void	bge_free_tx_ring(struct bge_softc *);
225 static int	bge_init_tx_ring(struct bge_softc *);
226 
227 static int	bge_chipinit(struct bge_softc *);
228 static int	bge_blockinit(struct bge_softc *);
229 static int	bge_setpowerstate(struct bge_softc *, int);
230 
231 static void	bge_reset(struct bge_softc *);
232 
233 #define BGE_DEBUG
234 #ifdef BGE_DEBUG
235 #define DPRINTF(x)	if (bgedebug) printf x
236 #define DPRINTFN(n,x)	if (bgedebug >= (n)) printf x
237 #define BGE_TSO_PRINTF(x)  do { if (bge_tso_debug) printf x ;} while (0)
238 int	bgedebug = 0;
239 int	bge_tso_debug = 0;
240 #else
241 #define DPRINTF(x)
242 #define DPRINTFN(n,x)
243 #define BGE_TSO_PRINTF(x)
244 #endif
245 
246 #ifdef BGE_EVENT_COUNTERS
247 #define	BGE_EVCNT_INCR(ev)	(ev).ev_count++
248 #define	BGE_EVCNT_ADD(ev, val)	(ev).ev_count += (val)
249 #define	BGE_EVCNT_UPD(ev, val)	(ev).ev_count = (val)
250 #else
251 #define	BGE_EVCNT_INCR(ev)	/* nothing */
252 #define	BGE_EVCNT_ADD(ev, val)	/* nothing */
253 #define	BGE_EVCNT_UPD(ev, val)	/* nothing */
254 #endif
255 
256 /* Various chip quirks. */
257 #define	BGE_QUIRK_LINK_STATE_BROKEN	0x00000001
258 #define	BGE_QUIRK_CSUM_BROKEN		0x00000002
259 #define	BGE_QUIRK_ONLY_PHY_1		0x00000004
260 #define	BGE_QUIRK_5700_SMALLDMA		0x00000008
261 #define	BGE_QUIRK_5700_PCIX_REG_BUG	0x00000010
262 #define	BGE_QUIRK_PRODUCER_BUG		0x00000020
263 #define	BGE_QUIRK_PCIX_DMA_ALIGN_BUG	0x00000040
264 #define	BGE_QUIRK_5705_CORE		0x00000080
265 #define	BGE_QUIRK_FEWER_MBUFS		0x00000100
266 
267 /*
268  * XXX: how to handle variants based on 5750 and derivatives:
269  * 5750 5751, 5721, possibly 5714, 5752, and 5708?, which
270  * in general behave like a 5705, except with additional quirks.
271  * This driver's current handling of the 5721 is wrong;
272  * how we map ASIC revision to "quirks" needs more thought.
273  * (defined here until the thought is done).
274  */
275 #define BGE_IS_5714_FAMILY(sc) \
276 	(BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5714_A0 || \
277 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5780 ||	\
278 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5714 )
279 
280 #define BGE_IS_5750_OR_BEYOND(sc)  \
281 	(BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5750 || \
282 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5752 || \
283 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5755 || \
284 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5787 || \
285 	 BGE_IS_5714_FAMILY(sc) )
286 
287 #define BGE_IS_5705_OR_BEYOND(sc)  \
288 	( ((sc)->bge_quirks & BGE_QUIRK_5705_CORE) || \
289 	  BGE_IS_5750_OR_BEYOND(sc) )
290 
291 
292 /* following bugs are common to bcm5700 rev B, all flavours */
293 #define BGE_QUIRK_5700_COMMON \
294 	(BGE_QUIRK_5700_SMALLDMA|BGE_QUIRK_PRODUCER_BUG)
295 
296 CFATTACH_DECL_NEW(bge, sizeof(struct bge_softc),
297     bge_probe, bge_attach, NULL, NULL);
298 
299 static u_int32_t
300 bge_readmem_ind(struct bge_softc *sc, int off)
301 {
302 	pcireg_t val;
303 
304 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MEMWIN_BASEADDR, off);
305 	val = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MEMWIN_DATA);
306 	return val;
307 }
308 
309 static void
310 bge_writemem_ind(struct bge_softc *sc, int off, int val)
311 {
312 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MEMWIN_BASEADDR, off);
313 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MEMWIN_DATA, val);
314 }
315 
316 #ifdef notdef
317 static u_int32_t
318 bge_readreg_ind(struct bge_softc *sc, int off)
319 {
320 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_REG_BASEADDR, off);
321 	return(pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_REG_DATA));
322 }
323 #endif
324 
325 static void
326 bge_writereg_ind(struct bge_softc *sc, int off, int val)
327 {
328 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_REG_BASEADDR, off);
329 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_REG_DATA, val);
330 }
331 
332 #ifdef notdef
333 static u_int8_t
334 bge_vpd_readbyte(struct bge_softc *sc, int addr)
335 {
336 	int i;
337 	u_int32_t val;
338 
339 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_VPD_ADDR, addr);
340 	for (i = 0; i < BGE_TIMEOUT * 10; i++) {
341 		DELAY(10);
342 		if (pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_VPD_ADDR) &
343 		    BGE_VPD_FLAG)
344 			break;
345 	}
346 
347 	if (i == BGE_TIMEOUT) {
348 		aprint_error_dev(sc->bge_dev, "VPD read timed out\n");
349 		return(0);
350 	}
351 
352 	val = pci_conf_read(sc->sc_pc, sc->sca_pcitag, BGE_PCI_VPD_DATA);
353 
354 	return((val >> ((addr % 4) * 8)) & 0xFF);
355 }
356 
357 static void
358 bge_vpd_read_res(struct bge_softc *sc, struct vpd_res *res, int addr)
359 {
360 	int i;
361 	u_int8_t *ptr;
362 
363 	ptr = (u_int8_t *)res;
364 	for (i = 0; i < sizeof(struct vpd_res); i++)
365 		ptr[i] = bge_vpd_readbyte(sc, i + addr);
366 }
367 
368 static void
369 bge_vpd_read(struct bge_softc *sc)
370 {
371 	int pos = 0, i;
372 	struct vpd_res res;
373 
374 	if (sc->bge_vpd_prodname != NULL)
375 		free(sc->bge_vpd_prodname, M_DEVBUF);
376 	if (sc->bge_vpd_readonly != NULL)
377 		free(sc->bge_vpd_readonly, M_DEVBUF);
378 	sc->bge_vpd_prodname = NULL;
379 	sc->bge_vpd_readonly = NULL;
380 
381 	bge_vpd_read_res(sc, &res, pos);
382 
383 	if (res.vr_id != VPD_RES_ID) {
384 		aprint_error_dev("bad VPD resource id: expected %x got %x\n",
385 		    VPD_RES_ID, res.vr_id);
386 		return;
387 	}
388 
389 	pos += sizeof(res);
390 	sc->bge_vpd_prodname = malloc(res.vr_len + 1, M_DEVBUF, M_NOWAIT);
391 	if (sc->bge_vpd_prodname == NULL)
392 		panic("bge_vpd_read");
393 	for (i = 0; i < res.vr_len; i++)
394 		sc->bge_vpd_prodname[i] = bge_vpd_readbyte(sc, i + pos);
395 	sc->bge_vpd_prodname[i] = '\0';
396 	pos += i;
397 
398 	bge_vpd_read_res(sc, &res, pos);
399 
400 	if (res.vr_id != VPD_RES_READ) {
401 		aprint_error_dev(sc->bge_dev,
402 		    "bad VPD resource id: expected %x got %x\n",
403 		    VPD_RES_READ, res.vr_id);
404 		return;
405 	}
406 
407 	pos += sizeof(res);
408 	sc->bge_vpd_readonly = malloc(res.vr_len, M_DEVBUF, M_NOWAIT);
409 	if (sc->bge_vpd_readonly == NULL)
410 		panic("bge_vpd_read");
411 	for (i = 0; i < res.vr_len + 1; i++)
412 		sc->bge_vpd_readonly[i] = bge_vpd_readbyte(sc, i + pos);
413 }
414 #endif
415 
416 /*
417  * Read a byte of data stored in the EEPROM at address 'addr.' The
418  * BCM570x supports both the traditional bitbang interface and an
419  * auto access interface for reading the EEPROM. We use the auto
420  * access method.
421  */
422 static u_int8_t
423 bge_eeprom_getbyte(struct bge_softc *sc, int addr, u_int8_t *dest)
424 {
425 	int i;
426 	u_int32_t byte = 0;
427 
428 	/*
429 	 * Enable use of auto EEPROM access so we can avoid
430 	 * having to use the bitbang method.
431 	 */
432 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_AUTO_EEPROM);
433 
434 	/* Reset the EEPROM, load the clock period. */
435 	CSR_WRITE_4(sc, BGE_EE_ADDR,
436 	    BGE_EEADDR_RESET|BGE_EEHALFCLK(BGE_HALFCLK_384SCL));
437 	DELAY(20);
438 
439 	/* Issue the read EEPROM command. */
440 	CSR_WRITE_4(sc, BGE_EE_ADDR, BGE_EE_READCMD | addr);
441 
442 	/* Wait for completion */
443 	for(i = 0; i < BGE_TIMEOUT * 10; i++) {
444 		DELAY(10);
445 		if (CSR_READ_4(sc, BGE_EE_ADDR) & BGE_EEADDR_DONE)
446 			break;
447 	}
448 
449 	if (i == BGE_TIMEOUT) {
450 		aprint_error_dev(sc->bge_dev, "eeprom read timed out\n");
451 		return(0);
452 	}
453 
454 	/* Get result. */
455 	byte = CSR_READ_4(sc, BGE_EE_DATA);
456 
457 	*dest = (byte >> ((addr % 4) * 8)) & 0xFF;
458 
459 	return(0);
460 }
461 
462 /*
463  * Read a sequence of bytes from the EEPROM.
464  */
465 static int
466 bge_read_eeprom(struct bge_softc *sc, void *destv, int off, int cnt)
467 {
468 	int err = 0, i;
469 	u_int8_t byte = 0;
470 	char *dest = destv;
471 
472 	for (i = 0; i < cnt; i++) {
473 		err = bge_eeprom_getbyte(sc, off + i, &byte);
474 		if (err)
475 			break;
476 		*(dest + i) = byte;
477 	}
478 
479 	return(err ? 1 : 0);
480 }
481 
482 static int
483 bge_miibus_readreg(device_t dev, int phy, int reg)
484 {
485 	struct bge_softc *sc = device_private(dev);
486 	u_int32_t val;
487 	u_int32_t saved_autopoll;
488 	int i;
489 
490 	/*
491 	 * Several chips with builtin PHYs will incorrectly answer to
492 	 * other PHY instances than the builtin PHY at id 1.
493 	 */
494 	if (phy != 1 && (sc->bge_quirks & BGE_QUIRK_ONLY_PHY_1))
495 		return(0);
496 
497 	/* Reading with autopolling on may trigger PCI errors */
498 	saved_autopoll = CSR_READ_4(sc, BGE_MI_MODE);
499 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
500 		CSR_WRITE_4(sc, BGE_MI_MODE,
501 		    saved_autopoll &~ BGE_MIMODE_AUTOPOLL);
502 		DELAY(40);
503 	}
504 
505 	CSR_WRITE_4(sc, BGE_MI_COMM, BGE_MICMD_READ|BGE_MICOMM_BUSY|
506 	    BGE_MIPHY(phy)|BGE_MIREG(reg));
507 
508 	for (i = 0; i < BGE_TIMEOUT; i++) {
509 		val = CSR_READ_4(sc, BGE_MI_COMM);
510 		if (!(val & BGE_MICOMM_BUSY))
511 			break;
512 		delay(10);
513 	}
514 
515 	if (i == BGE_TIMEOUT) {
516 		aprint_error_dev(sc->bge_dev, "PHY read timed out\n");
517 		val = 0;
518 		goto done;
519 	}
520 
521 	val = CSR_READ_4(sc, BGE_MI_COMM);
522 
523 done:
524 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
525 		CSR_WRITE_4(sc, BGE_MI_MODE, saved_autopoll);
526 		DELAY(40);
527 	}
528 
529 	if (val & BGE_MICOMM_READFAIL)
530 		return(0);
531 
532 	return(val & 0xFFFF);
533 }
534 
535 static void
536 bge_miibus_writereg(device_t dev, int phy, int reg, int val)
537 {
538 	struct bge_softc *sc = device_private(dev);
539 	u_int32_t saved_autopoll;
540 	int i;
541 
542 	/* Touching the PHY while autopolling is on may trigger PCI errors */
543 	saved_autopoll = CSR_READ_4(sc, BGE_MI_MODE);
544 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
545 		delay(40);
546 		CSR_WRITE_4(sc, BGE_MI_MODE,
547 		    saved_autopoll & (~BGE_MIMODE_AUTOPOLL));
548 		delay(10); /* 40 usec is supposed to be adequate */
549 	}
550 
551 	CSR_WRITE_4(sc, BGE_MI_COMM, BGE_MICMD_WRITE|BGE_MICOMM_BUSY|
552 	    BGE_MIPHY(phy)|BGE_MIREG(reg)|val);
553 
554 	for (i = 0; i < BGE_TIMEOUT; i++) {
555 		if (!(CSR_READ_4(sc, BGE_MI_COMM) & BGE_MICOMM_BUSY))
556 			break;
557 		delay(10);
558 	}
559 
560 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
561 		CSR_WRITE_4(sc, BGE_MI_MODE, saved_autopoll);
562 		delay(40);
563 	}
564 
565 	if (i == BGE_TIMEOUT)
566 		aprint_error_dev(sc->bge_dev, "PHY read timed out\n");
567 }
568 
569 static void
570 bge_miibus_statchg(device_t dev)
571 {
572 	struct bge_softc *sc = device_private(dev);
573 	struct mii_data *mii = &sc->bge_mii;
574 
575 	/*
576 	 * Get flow control negotiation result.
577 	 */
578 	if (IFM_SUBTYPE(mii->mii_media.ifm_cur->ifm_media) == IFM_AUTO &&
579 	    (mii->mii_media_active & IFM_ETH_FMASK) != sc->bge_flowflags) {
580 		sc->bge_flowflags = mii->mii_media_active & IFM_ETH_FMASK;
581 		mii->mii_media_active &= ~IFM_ETH_FMASK;
582 	}
583 
584 	BGE_CLRBIT(sc, BGE_MAC_MODE, BGE_MACMODE_PORTMODE);
585 	if (IFM_SUBTYPE(mii->mii_media_active) == IFM_1000_T) {
586 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_PORTMODE_GMII);
587 	} else {
588 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_PORTMODE_MII);
589 	}
590 
591 	if ((mii->mii_media_active & IFM_GMASK) == IFM_FDX) {
592 		BGE_CLRBIT(sc, BGE_MAC_MODE, BGE_MACMODE_HALF_DUPLEX);
593 	} else {
594 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_MACMODE_HALF_DUPLEX);
595 	}
596 
597 	/*
598 	 * 802.3x flow control
599 	 */
600 	if (sc->bge_flowflags & IFM_ETH_RXPAUSE) {
601 		BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_FLOWCTL_ENABLE);
602 	} else {
603 		BGE_CLRBIT(sc, BGE_RX_MODE, BGE_RXMODE_FLOWCTL_ENABLE);
604 	}
605 	if (sc->bge_flowflags & IFM_ETH_TXPAUSE) {
606 		BGE_SETBIT(sc, BGE_TX_MODE, BGE_TXMODE_FLOWCTL_ENABLE);
607 	} else {
608 		BGE_CLRBIT(sc, BGE_TX_MODE, BGE_TXMODE_FLOWCTL_ENABLE);
609 	}
610 }
611 
612 /*
613  * Update rx threshold levels to values in a particular slot
614  * of the interrupt-mitigation table bge_rx_threshes.
615  */
616 static void
617 bge_set_thresh(struct ifnet *ifp, int lvl)
618 {
619 	struct bge_softc *sc = ifp->if_softc;
620 	int s;
621 
622 	/* For now, just save the new Rx-intr thresholds and record
623 	 * that a threshold update is pending.  Updating the hardware
624 	 * registers here (even at splhigh()) is observed to
625 	 * occasionaly cause glitches where Rx-interrupts are not
626 	 * honoured for up to 10 seconds. jonathan@NetBSD.org, 2003-04-05
627 	 */
628 	s = splnet();
629 	sc->bge_rx_coal_ticks = bge_rx_threshes[lvl].rx_ticks;
630 	sc->bge_rx_max_coal_bds = bge_rx_threshes[lvl].rx_max_bds;
631 	sc->bge_pending_rxintr_change = 1;
632 	splx(s);
633 
634 	 return;
635 }
636 
637 
638 /*
639  * Update Rx thresholds of all bge devices
640  */
641 static void
642 bge_update_all_threshes(int lvl)
643 {
644 	struct ifnet *ifp;
645 	const char * const namebuf = "bge";
646 	int namelen;
647 
648 	if (lvl < 0)
649 		lvl = 0;
650 	else if( lvl >= NBGE_RX_THRESH)
651 		lvl = NBGE_RX_THRESH - 1;
652 
653 	namelen = strlen(namebuf);
654 	/*
655 	 * Now search all the interfaces for this name/number
656 	 */
657 	IFNET_FOREACH(ifp) {
658 		if (strncmp(ifp->if_xname, namebuf, namelen) != 0)
659 		      continue;
660 		/* We got a match: update if doing auto-threshold-tuning */
661 		if (bge_auto_thresh)
662 			bge_set_thresh(ifp, lvl);
663 	}
664 }
665 
666 /*
667  * Handle events that have triggered interrupts.
668  */
669 static void
670 bge_handle_events(struct bge_softc *sc)
671 {
672 
673 	return;
674 }
675 
676 /*
677  * Memory management for jumbo frames.
678  */
679 
680 static int
681 bge_alloc_jumbo_mem(struct bge_softc *sc)
682 {
683 	char *ptr, *kva;
684 	bus_dma_segment_t	seg;
685 	int		i, rseg, state, error;
686 	struct bge_jpool_entry   *entry;
687 
688 	state = error = 0;
689 
690 	/* Grab a big chunk o' storage. */
691 	if (bus_dmamem_alloc(sc->bge_dmatag, BGE_JMEM, PAGE_SIZE, 0,
692 	     &seg, 1, &rseg, BUS_DMA_NOWAIT)) {
693 		aprint_error_dev(sc->bge_dev, "can't alloc rx buffers\n");
694 		return ENOBUFS;
695 	}
696 
697 	state = 1;
698 	if (bus_dmamem_map(sc->bge_dmatag, &seg, rseg, BGE_JMEM, (void **)&kva,
699 	    BUS_DMA_NOWAIT)) {
700 		aprint_error_dev(sc->bge_dev,
701 		    "can't map DMA buffers (%d bytes)\n", (int)BGE_JMEM);
702 		error = ENOBUFS;
703 		goto out;
704 	}
705 
706 	state = 2;
707 	if (bus_dmamap_create(sc->bge_dmatag, BGE_JMEM, 1, BGE_JMEM, 0,
708 	    BUS_DMA_NOWAIT, &sc->bge_cdata.bge_rx_jumbo_map)) {
709 		aprint_error_dev(sc->bge_dev, "can't create DMA map\n");
710 		error = ENOBUFS;
711 		goto out;
712 	}
713 
714 	state = 3;
715 	if (bus_dmamap_load(sc->bge_dmatag, sc->bge_cdata.bge_rx_jumbo_map,
716 	    kva, BGE_JMEM, NULL, BUS_DMA_NOWAIT)) {
717 		aprint_error_dev(sc->bge_dev, "can't load DMA map\n");
718 		error = ENOBUFS;
719 		goto out;
720 	}
721 
722 	state = 4;
723 	sc->bge_cdata.bge_jumbo_buf = (void *)kva;
724 	DPRINTFN(1,("bge_jumbo_buf = %p\n", sc->bge_cdata.bge_jumbo_buf));
725 
726 	SLIST_INIT(&sc->bge_jfree_listhead);
727 	SLIST_INIT(&sc->bge_jinuse_listhead);
728 
729 	/*
730 	 * Now divide it up into 9K pieces and save the addresses
731 	 * in an array.
732 	 */
733 	ptr = sc->bge_cdata.bge_jumbo_buf;
734 	for (i = 0; i < BGE_JSLOTS; i++) {
735 		sc->bge_cdata.bge_jslots[i] = ptr;
736 		ptr += BGE_JLEN;
737 		entry = malloc(sizeof(struct bge_jpool_entry),
738 		    M_DEVBUF, M_NOWAIT);
739 		if (entry == NULL) {
740 			aprint_error_dev(sc->bge_dev,
741 			    "no memory for jumbo buffer queue!\n");
742 			error = ENOBUFS;
743 			goto out;
744 		}
745 		entry->slot = i;
746 		SLIST_INSERT_HEAD(&sc->bge_jfree_listhead,
747 				 entry, jpool_entries);
748 	}
749 out:
750 	if (error != 0) {
751 		switch (state) {
752 		case 4:
753 			bus_dmamap_unload(sc->bge_dmatag,
754 			    sc->bge_cdata.bge_rx_jumbo_map);
755 		case 3:
756 			bus_dmamap_destroy(sc->bge_dmatag,
757 			    sc->bge_cdata.bge_rx_jumbo_map);
758 		case 2:
759 			bus_dmamem_unmap(sc->bge_dmatag, kva, BGE_JMEM);
760 		case 1:
761 			bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
762 			break;
763 		default:
764 			break;
765 		}
766 	}
767 
768 	return error;
769 }
770 
771 /*
772  * Allocate a jumbo buffer.
773  */
774 static void *
775 bge_jalloc(struct bge_softc *sc)
776 {
777 	struct bge_jpool_entry   *entry;
778 
779 	entry = SLIST_FIRST(&sc->bge_jfree_listhead);
780 
781 	if (entry == NULL) {
782 		aprint_error_dev(sc->bge_dev, "no free jumbo buffers\n");
783 		return(NULL);
784 	}
785 
786 	SLIST_REMOVE_HEAD(&sc->bge_jfree_listhead, jpool_entries);
787 	SLIST_INSERT_HEAD(&sc->bge_jinuse_listhead, entry, jpool_entries);
788 	return(sc->bge_cdata.bge_jslots[entry->slot]);
789 }
790 
791 /*
792  * Release a jumbo buffer.
793  */
794 static void
795 bge_jfree(struct mbuf *m, void *buf, size_t size, void *arg)
796 {
797 	struct bge_jpool_entry *entry;
798 	struct bge_softc *sc;
799 	int i, s;
800 
801 	/* Extract the softc struct pointer. */
802 	sc = (struct bge_softc *)arg;
803 
804 	if (sc == NULL)
805 		panic("bge_jfree: can't find softc pointer!");
806 
807 	/* calculate the slot this buffer belongs to */
808 
809 	i = ((char *)buf
810 	     - (char *)sc->bge_cdata.bge_jumbo_buf) / BGE_JLEN;
811 
812 	if ((i < 0) || (i >= BGE_JSLOTS))
813 		panic("bge_jfree: asked to free buffer that we don't manage!");
814 
815 	s = splvm();
816 	entry = SLIST_FIRST(&sc->bge_jinuse_listhead);
817 	if (entry == NULL)
818 		panic("bge_jfree: buffer not in use!");
819 	entry->slot = i;
820 	SLIST_REMOVE_HEAD(&sc->bge_jinuse_listhead, jpool_entries);
821 	SLIST_INSERT_HEAD(&sc->bge_jfree_listhead, entry, jpool_entries);
822 
823 	if (__predict_true(m != NULL))
824   		pool_cache_put(mb_cache, m);
825 	splx(s);
826 }
827 
828 
829 /*
830  * Intialize a standard receive ring descriptor.
831  */
832 static int
833 bge_newbuf_std(struct bge_softc *sc, int i, struct mbuf *m, bus_dmamap_t dmamap)
834 {
835 	struct mbuf		*m_new = NULL;
836 	struct bge_rx_bd	*r;
837 	int			error;
838 
839 	if (dmamap == NULL) {
840 		error = bus_dmamap_create(sc->bge_dmatag, MCLBYTES, 1,
841 		    MCLBYTES, 0, BUS_DMA_NOWAIT, &dmamap);
842 		if (error != 0)
843 			return error;
844 	}
845 
846 	sc->bge_cdata.bge_rx_std_map[i] = dmamap;
847 
848 	if (m == NULL) {
849 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
850 		if (m_new == NULL) {
851 			return(ENOBUFS);
852 		}
853 
854 		MCLGET(m_new, M_DONTWAIT);
855 		if (!(m_new->m_flags & M_EXT)) {
856 			m_freem(m_new);
857 			return(ENOBUFS);
858 		}
859 		m_new->m_len = m_new->m_pkthdr.len = MCLBYTES;
860 
861 	} else {
862 		m_new = m;
863 		m_new->m_len = m_new->m_pkthdr.len = MCLBYTES;
864 		m_new->m_data = m_new->m_ext.ext_buf;
865 	}
866 	if (!sc->bge_rx_alignment_bug)
867 	    m_adj(m_new, ETHER_ALIGN);
868 	if (bus_dmamap_load_mbuf(sc->bge_dmatag, dmamap, m_new,
869 	    BUS_DMA_READ|BUS_DMA_NOWAIT))
870 		return(ENOBUFS);
871 	bus_dmamap_sync(sc->bge_dmatag, dmamap, 0, dmamap->dm_mapsize,
872 	    BUS_DMASYNC_PREREAD);
873 
874 	sc->bge_cdata.bge_rx_std_chain[i] = m_new;
875 	r = &sc->bge_rdata->bge_rx_std_ring[i];
876 	bge_set_hostaddr(&r->bge_addr,
877 	    dmamap->dm_segs[0].ds_addr);
878 	r->bge_flags = BGE_RXBDFLAG_END;
879 	r->bge_len = m_new->m_len;
880 	r->bge_idx = i;
881 
882 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
883 	    offsetof(struct bge_ring_data, bge_rx_std_ring) +
884 		i * sizeof (struct bge_rx_bd),
885 	    sizeof (struct bge_rx_bd),
886 	    BUS_DMASYNC_PREWRITE|BUS_DMASYNC_PREREAD);
887 
888 	return(0);
889 }
890 
891 /*
892  * Initialize a jumbo receive ring descriptor. This allocates
893  * a jumbo buffer from the pool managed internally by the driver.
894  */
895 static int
896 bge_newbuf_jumbo(struct bge_softc *sc, int i, struct mbuf *m)
897 {
898 	struct mbuf *m_new = NULL;
899 	struct bge_rx_bd *r;
900 	void *buf = NULL;
901 
902 	if (m == NULL) {
903 
904 		/* Allocate the mbuf. */
905 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
906 		if (m_new == NULL) {
907 			return(ENOBUFS);
908 		}
909 
910 		/* Allocate the jumbo buffer */
911 		buf = bge_jalloc(sc);
912 		if (buf == NULL) {
913 			m_freem(m_new);
914 			aprint_error_dev(sc->bge_dev,
915 			    "jumbo allocation failed -- packet dropped!\n");
916 			return(ENOBUFS);
917 		}
918 
919 		/* Attach the buffer to the mbuf. */
920 		m_new->m_len = m_new->m_pkthdr.len = BGE_JUMBO_FRAMELEN;
921 		MEXTADD(m_new, buf, BGE_JUMBO_FRAMELEN, M_DEVBUF,
922 		    bge_jfree, sc);
923 		m_new->m_flags |= M_EXT_RW;
924 	} else {
925 		m_new = m;
926 		buf = m_new->m_data = m_new->m_ext.ext_buf;
927 		m_new->m_ext.ext_size = BGE_JUMBO_FRAMELEN;
928 	}
929 	if (!sc->bge_rx_alignment_bug)
930 	    m_adj(m_new, ETHER_ALIGN);
931 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_cdata.bge_rx_jumbo_map,
932 	    mtod(m_new, char *) - (char *)sc->bge_cdata.bge_jumbo_buf, BGE_JLEN,
933 	    BUS_DMASYNC_PREREAD);
934 	/* Set up the descriptor. */
935 	r = &sc->bge_rdata->bge_rx_jumbo_ring[i];
936 	sc->bge_cdata.bge_rx_jumbo_chain[i] = m_new;
937 	bge_set_hostaddr(&r->bge_addr, BGE_JUMBO_DMA_ADDR(sc, m_new));
938 	r->bge_flags = BGE_RXBDFLAG_END|BGE_RXBDFLAG_JUMBO_RING;
939 	r->bge_len = m_new->m_len;
940 	r->bge_idx = i;
941 
942 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
943 	    offsetof(struct bge_ring_data, bge_rx_jumbo_ring) +
944 		i * sizeof (struct bge_rx_bd),
945 	    sizeof (struct bge_rx_bd),
946 	    BUS_DMASYNC_PREWRITE|BUS_DMASYNC_PREREAD);
947 
948 	return(0);
949 }
950 
951 /*
952  * The standard receive ring has 512 entries in it. At 2K per mbuf cluster,
953  * that's 1MB or memory, which is a lot. For now, we fill only the first
954  * 256 ring entries and hope that our CPU is fast enough to keep up with
955  * the NIC.
956  */
957 static int
958 bge_init_rx_ring_std(struct bge_softc *sc)
959 {
960 	int i;
961 
962 	if (sc->bge_flags & BGE_RXRING_VALID)
963 		return 0;
964 
965 	for (i = 0; i < BGE_SSLOTS; i++) {
966 		if (bge_newbuf_std(sc, i, NULL, 0) == ENOBUFS)
967 			return(ENOBUFS);
968 	}
969 
970 	sc->bge_std = i - 1;
971 	CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, sc->bge_std);
972 
973 	sc->bge_flags |= BGE_RXRING_VALID;
974 
975 	return(0);
976 }
977 
978 static void
979 bge_free_rx_ring_std(struct bge_softc *sc)
980 {
981 	int i;
982 
983 	if (!(sc->bge_flags & BGE_RXRING_VALID))
984 		return;
985 
986 	for (i = 0; i < BGE_STD_RX_RING_CNT; i++) {
987 		if (sc->bge_cdata.bge_rx_std_chain[i] != NULL) {
988 			m_freem(sc->bge_cdata.bge_rx_std_chain[i]);
989 			sc->bge_cdata.bge_rx_std_chain[i] = NULL;
990 			bus_dmamap_destroy(sc->bge_dmatag,
991 			    sc->bge_cdata.bge_rx_std_map[i]);
992 		}
993 		memset((char *)&sc->bge_rdata->bge_rx_std_ring[i], 0,
994 		    sizeof(struct bge_rx_bd));
995 	}
996 
997 	sc->bge_flags &= ~BGE_RXRING_VALID;
998 }
999 
1000 static int
1001 bge_init_rx_ring_jumbo(struct bge_softc *sc)
1002 {
1003 	int i;
1004 	volatile struct bge_rcb *rcb;
1005 
1006 	if (sc->bge_flags & BGE_JUMBO_RXRING_VALID)
1007 		return 0;
1008 
1009 	for (i = 0; i < BGE_JUMBO_RX_RING_CNT; i++) {
1010 		if (bge_newbuf_jumbo(sc, i, NULL) == ENOBUFS)
1011 			return(ENOBUFS);
1012 	};
1013 
1014 	sc->bge_jumbo = i - 1;
1015 	sc->bge_flags |= BGE_JUMBO_RXRING_VALID;
1016 
1017 	rcb = &sc->bge_rdata->bge_info.bge_jumbo_rx_rcb;
1018 	rcb->bge_maxlen_flags = 0;
1019 	CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_MAXLEN_FLAGS, rcb->bge_maxlen_flags);
1020 
1021 	CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, sc->bge_jumbo);
1022 
1023 	return(0);
1024 }
1025 
1026 static void
1027 bge_free_rx_ring_jumbo(struct bge_softc *sc)
1028 {
1029 	int i;
1030 
1031 	if (!(sc->bge_flags & BGE_JUMBO_RXRING_VALID))
1032 		return;
1033 
1034 	for (i = 0; i < BGE_JUMBO_RX_RING_CNT; i++) {
1035 		if (sc->bge_cdata.bge_rx_jumbo_chain[i] != NULL) {
1036 			m_freem(sc->bge_cdata.bge_rx_jumbo_chain[i]);
1037 			sc->bge_cdata.bge_rx_jumbo_chain[i] = NULL;
1038 		}
1039 		memset((char *)&sc->bge_rdata->bge_rx_jumbo_ring[i], 0,
1040 		    sizeof(struct bge_rx_bd));
1041 	}
1042 
1043 	sc->bge_flags &= ~BGE_JUMBO_RXRING_VALID;
1044 }
1045 
1046 static void
1047 bge_free_tx_ring(struct bge_softc *sc)
1048 {
1049 	int i, freed;
1050 	struct txdmamap_pool_entry *dma;
1051 
1052 	if (!(sc->bge_flags & BGE_TXRING_VALID))
1053 		return;
1054 
1055 	freed = 0;
1056 
1057 	for (i = 0; i < BGE_TX_RING_CNT; i++) {
1058 		if (sc->bge_cdata.bge_tx_chain[i] != NULL) {
1059 			freed++;
1060 			m_freem(sc->bge_cdata.bge_tx_chain[i]);
1061 			sc->bge_cdata.bge_tx_chain[i] = NULL;
1062 			SLIST_INSERT_HEAD(&sc->txdma_list, sc->txdma[i],
1063 					    link);
1064 			sc->txdma[i] = 0;
1065 		}
1066 		memset((char *)&sc->bge_rdata->bge_tx_ring[i], 0,
1067 		    sizeof(struct bge_tx_bd));
1068 	}
1069 
1070 	while ((dma = SLIST_FIRST(&sc->txdma_list))) {
1071 		SLIST_REMOVE_HEAD(&sc->txdma_list, link);
1072 		bus_dmamap_destroy(sc->bge_dmatag, dma->dmamap);
1073 		free(dma, M_DEVBUF);
1074 	}
1075 
1076 	sc->bge_flags &= ~BGE_TXRING_VALID;
1077 }
1078 
1079 static int
1080 bge_init_tx_ring(struct bge_softc *sc)
1081 {
1082 	int i;
1083 	bus_dmamap_t dmamap;
1084 	struct txdmamap_pool_entry *dma;
1085 
1086 	if (sc->bge_flags & BGE_TXRING_VALID)
1087 		return 0;
1088 
1089 	sc->bge_txcnt = 0;
1090 	sc->bge_tx_saved_considx = 0;
1091 
1092 	/* Initialize transmit producer index for host-memory send ring. */
1093 	sc->bge_tx_prodidx = 0;
1094 	CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, sc->bge_tx_prodidx);
1095 	if (sc->bge_quirks & BGE_QUIRK_PRODUCER_BUG)	/* 5700 b2 errata */
1096 		CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, sc->bge_tx_prodidx);
1097 
1098 	/* NIC-memory send ring  not used; initialize to zero. */
1099 	CSR_WRITE_4(sc, BGE_MBX_TX_NIC_PROD0_LO, 0);
1100 	if (sc->bge_quirks & BGE_QUIRK_PRODUCER_BUG)	/* 5700 b2 errata */
1101 		CSR_WRITE_4(sc, BGE_MBX_TX_NIC_PROD0_LO, 0);
1102 
1103 	SLIST_INIT(&sc->txdma_list);
1104 	for (i = 0; i < BGE_RSLOTS; i++) {
1105 		if (bus_dmamap_create(sc->bge_dmatag, BGE_TXDMA_MAX,
1106 		    BGE_NTXSEG, ETHER_MAX_LEN_JUMBO, 0, BUS_DMA_NOWAIT,
1107 		    &dmamap))
1108 			return(ENOBUFS);
1109 		if (dmamap == NULL)
1110 			panic("dmamap NULL in bge_init_tx_ring");
1111 		dma = malloc(sizeof(*dma), M_DEVBUF, M_NOWAIT);
1112 		if (dma == NULL) {
1113 			aprint_error_dev(sc->bge_dev,
1114 			    "can't alloc txdmamap_pool_entry\n");
1115 			bus_dmamap_destroy(sc->bge_dmatag, dmamap);
1116 			return (ENOMEM);
1117 		}
1118 		dma->dmamap = dmamap;
1119 		SLIST_INSERT_HEAD(&sc->txdma_list, dma, link);
1120 	}
1121 
1122 	sc->bge_flags |= BGE_TXRING_VALID;
1123 
1124 	return(0);
1125 }
1126 
1127 static void
1128 bge_setmulti(struct bge_softc *sc)
1129 {
1130 	struct ethercom		*ac = &sc->ethercom;
1131 	struct ifnet		*ifp = &ac->ec_if;
1132 	struct ether_multi	*enm;
1133 	struct ether_multistep  step;
1134 	u_int32_t		hashes[4] = { 0, 0, 0, 0 };
1135 	u_int32_t		h;
1136 	int			i;
1137 
1138 	if (ifp->if_flags & IFF_PROMISC)
1139 		goto allmulti;
1140 
1141 	/* Now program new ones. */
1142 	ETHER_FIRST_MULTI(step, ac, enm);
1143 	while (enm != NULL) {
1144 		if (memcmp(enm->enm_addrlo, enm->enm_addrhi, ETHER_ADDR_LEN)) {
1145 			/*
1146 			 * We must listen to a range of multicast addresses.
1147 			 * For now, just accept all multicasts, rather than
1148 			 * trying to set only those filter bits needed to match
1149 			 * the range.  (At this time, the only use of address
1150 			 * ranges is for IP multicast routing, for which the
1151 			 * range is big enough to require all bits set.)
1152 			 */
1153 			goto allmulti;
1154 		}
1155 
1156 		h = ether_crc32_le(enm->enm_addrlo, ETHER_ADDR_LEN);
1157 
1158 		/* Just want the 7 least-significant bits. */
1159 		h &= 0x7f;
1160 
1161 		hashes[(h & 0x60) >> 5] |= 1 << (h & 0x1F);
1162 		ETHER_NEXT_MULTI(step, enm);
1163 	}
1164 
1165 	ifp->if_flags &= ~IFF_ALLMULTI;
1166 	goto setit;
1167 
1168  allmulti:
1169 	ifp->if_flags |= IFF_ALLMULTI;
1170 	hashes[0] = hashes[1] = hashes[2] = hashes[3] = 0xffffffff;
1171 
1172  setit:
1173 	for (i = 0; i < 4; i++)
1174 		CSR_WRITE_4(sc, BGE_MAR0 + (i * 4), hashes[i]);
1175 }
1176 
1177 const int bge_swapbits[] = {
1178 	0,
1179 	BGE_MODECTL_BYTESWAP_DATA,
1180 	BGE_MODECTL_WORDSWAP_DATA,
1181 	BGE_MODECTL_BYTESWAP_NONFRAME,
1182 	BGE_MODECTL_WORDSWAP_NONFRAME,
1183 
1184 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA,
1185 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME,
1186 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_NONFRAME,
1187 
1188 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME,
1189 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_WORDSWAP_NONFRAME,
1190 
1191 	BGE_MODECTL_BYTESWAP_NONFRAME|BGE_MODECTL_WORDSWAP_NONFRAME,
1192 
1193 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
1194 	    BGE_MODECTL_BYTESWAP_NONFRAME,
1195 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
1196 	    BGE_MODECTL_WORDSWAP_NONFRAME,
1197 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME|
1198 	    BGE_MODECTL_WORDSWAP_NONFRAME,
1199 	BGE_MODECTL_WORDSWAP_DATA|BGE_MODECTL_BYTESWAP_NONFRAME|
1200 	    BGE_MODECTL_WORDSWAP_NONFRAME,
1201 
1202 	BGE_MODECTL_BYTESWAP_DATA|BGE_MODECTL_WORDSWAP_DATA|
1203 	    BGE_MODECTL_BYTESWAP_NONFRAME|BGE_MODECTL_WORDSWAP_NONFRAME,
1204 };
1205 
1206 int bge_swapindex = 0;
1207 
1208 /*
1209  * Do endian, PCI and DMA initialization. Also check the on-board ROM
1210  * self-test results.
1211  */
1212 static int
1213 bge_chipinit(struct bge_softc *sc)
1214 {
1215 	u_int32_t		cachesize;
1216 	int			i;
1217 	u_int32_t		dma_rw_ctl;
1218 
1219 
1220 	/* Set endianness before we access any non-PCI registers. */
1221 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MISC_CTL,
1222 	    BGE_INIT);
1223 
1224 	/* Set power state to D0. */
1225 	bge_setpowerstate(sc, 0);
1226 
1227 	/*
1228 	 * Check the 'ROM failed' bit on the RX CPU to see if
1229 	 * self-tests passed.
1230 	 */
1231 	if (CSR_READ_4(sc, BGE_RXCPU_MODE) & BGE_RXCPUMODE_ROMFAIL) {
1232 		aprint_error_dev(sc->bge_dev,
1233 		    "RX CPU self-diagnostics failed!\n");
1234 		return(ENODEV);
1235 	}
1236 
1237 	/* Clear the MAC control register */
1238 	CSR_WRITE_4(sc, BGE_MAC_MODE, 0);
1239 
1240 	/*
1241 	 * Clear the MAC statistics block in the NIC's
1242 	 * internal memory.
1243 	 */
1244 	for (i = BGE_STATS_BLOCK;
1245 	    i < BGE_STATS_BLOCK_END + 1; i += sizeof(u_int32_t))
1246 		BGE_MEMWIN_WRITE(sc->sc_pc, sc->sc_pcitag, i, 0);
1247 
1248 	for (i = BGE_STATUS_BLOCK;
1249 	    i < BGE_STATUS_BLOCK_END + 1; i += sizeof(u_int32_t))
1250 		BGE_MEMWIN_WRITE(sc->sc_pc, sc->sc_pcitag, i, 0);
1251 
1252 	/* Set up the PCI DMA control register. */
1253 	if (sc->bge_pcie) {
1254 	  u_int32_t device_ctl;
1255 
1256 		/* From FreeBSD */
1257 		DPRINTFN(4, ("(%s: PCI-Express DMA setting)\n",
1258 		    device_xname(sc->bge_dev)));
1259 		dma_rw_ctl = (BGE_PCI_READ_CMD | BGE_PCI_WRITE_CMD |
1260 		    (0xf << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
1261 		    (0x2 << BGE_PCIDMARWCTL_WR_WAT_SHIFT));
1262 
1263 		/* jonathan: alternative from Linux driver */
1264 #define DMA_CTRL_WRITE_PCIE_H20MARK_128         0x00180000
1265 #define DMA_CTRL_WRITE_PCIE_H20MARK_256         0x00380000
1266 
1267 		dma_rw_ctl =   0x76000000; /* XXX XXX XXX */;
1268 		device_ctl = pci_conf_read(sc->sc_pc, sc->sc_pcitag,
1269 					   BGE_PCI_CONF_DEV_CTRL);
1270 		aprint_debug_dev(sc->bge_dev, "pcie mode=0x%x\n", device_ctl);
1271 
1272 		if ((device_ctl & 0x00e0) && 0) {
1273 			/*
1274 			 * XXX jonathan@NetBSD.org:
1275 			 * This clause is exactly what the Broadcom-supplied
1276 			 * Linux does; but given overall register programming
1277 			 * by if_bge(4), this larger DMA-write watermark
1278 			 * value causes bcm5721 chips to totally wedge.
1279 			 */
1280 			dma_rw_ctl |= BGE_PCIDMA_RWCTL_PCIE_WRITE_WATRMARK_256;
1281 		} else {
1282 			dma_rw_ctl |= BGE_PCIDMA_RWCTL_PCIE_WRITE_WATRMARK_128;
1283 		}
1284 	} else if (pci_conf_read(sc->sc_pc, sc->sc_pcitag,BGE_PCI_PCISTATE) &
1285 	    BGE_PCISTATE_PCI_BUSMODE) {
1286 		/* Conventional PCI bus */
1287 	  	DPRINTFN(4, ("(%s: PCI 2.2 DMA setting)\n",
1288 		    device_xname(sc->bge_dev)));
1289 		dma_rw_ctl = (BGE_PCI_READ_CMD | BGE_PCI_WRITE_CMD |
1290 		   (0x7 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
1291 		   (0x7 << BGE_PCIDMARWCTL_WR_WAT_SHIFT));
1292 		if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1293 			dma_rw_ctl |= 0x0F;
1294 		}
1295 	} else {
1296 	  	DPRINTFN(4, ("(:%s: PCI-X DMA setting)\n",
1297 		    device_xname(sc->bge_dev)));
1298 		/* PCI-X bus */
1299 		dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD |
1300 		    (0x3 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
1301 		    (0x3 << BGE_PCIDMARWCTL_WR_WAT_SHIFT) |
1302 		    (0x0F);
1303 		/*
1304 		 * 5703 and 5704 need ONEDMA_AT_ONCE as a workaround
1305 		 * for hardware bugs, which means we should also clear
1306 		 * the low-order MINDMA bits.  In addition, the 5704
1307 		 * uses a different encoding of read/write watermarks.
1308 		 */
1309 		if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5704) {
1310 			dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD |
1311 			  /* should be 0x1f0000 */
1312 			  (0x7 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
1313 			  (0x3 << BGE_PCIDMARWCTL_WR_WAT_SHIFT);
1314 			dma_rw_ctl |= BGE_PCIDMARWCTL_ONEDMA_ATONCE;
1315 		}
1316 		else if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5703) {
1317 			dma_rw_ctl &=  0xfffffff0;
1318 			dma_rw_ctl |= BGE_PCIDMARWCTL_ONEDMA_ATONCE;
1319 		}
1320 		else if (BGE_IS_5714_FAMILY(sc)) {
1321 			dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD;
1322 			dma_rw_ctl &= ~BGE_PCIDMARWCTL_ONEDMA_ATONCE; /* XXX */
1323 			/* XXX magic values, Broadcom-supplied Linux driver */
1324 			if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5780)
1325 				dma_rw_ctl |= (1 << 20) | (1 << 18) |
1326 				  BGE_PCIDMARWCTL_ONEDMA_ATONCE;
1327 			else
1328 				dma_rw_ctl |= (1<<20) | (1<<18) | (1 << 15);
1329 		}
1330 	}
1331 
1332 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL, dma_rw_ctl);
1333 
1334 	/*
1335 	 * Set up general mode register.
1336 	 */
1337 	CSR_WRITE_4(sc, BGE_MODE_CTL, BGE_DMA_SWAP_OPTIONS|
1338 		    BGE_MODECTL_MAC_ATTN_INTR|BGE_MODECTL_HOST_SEND_BDS|
1339 		    BGE_MODECTL_TX_NO_PHDR_CSUM|BGE_MODECTL_RX_NO_PHDR_CSUM);
1340 
1341 	/* Get cache line size. */
1342 	cachesize = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CACHESZ);
1343 
1344 	/*
1345 	 * Avoid violating PCI spec on certain chip revs.
1346 	 */
1347 	if (pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CMD) &
1348 	    PCIM_CMD_MWIEN) {
1349 		switch(cachesize) {
1350 		case 1:
1351 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1352 				   BGE_PCI_WRITE_BNDRY_16BYTES);
1353 			break;
1354 		case 2:
1355 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1356 				   BGE_PCI_WRITE_BNDRY_32BYTES);
1357 			break;
1358 		case 4:
1359 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1360 				   BGE_PCI_WRITE_BNDRY_64BYTES);
1361 			break;
1362 		case 8:
1363 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1364 				   BGE_PCI_WRITE_BNDRY_128BYTES);
1365 			break;
1366 		case 16:
1367 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1368 				   BGE_PCI_WRITE_BNDRY_256BYTES);
1369 			break;
1370 		case 32:
1371 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1372 				   BGE_PCI_WRITE_BNDRY_512BYTES);
1373 			break;
1374 		case 64:
1375 			PCI_SETBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_DMA_RW_CTL,
1376 				   BGE_PCI_WRITE_BNDRY_1024BYTES);
1377 			break;
1378 		default:
1379 		/* Disable PCI memory write and invalidate. */
1380 #if 0
1381 			if (bootverbose)
1382 				aprint_error_dev(sc->bge_dev,
1383 				    "cache line size %d not supported "
1384 				    "disabling PCI MWI\n",
1385 #endif
1386 			PCI_CLRBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CMD,
1387 			    PCIM_CMD_MWIEN);
1388 			break;
1389 		}
1390 	}
1391 
1392 	/*
1393 	 * Disable memory write invalidate.  Apparently it is not supported
1394 	 * properly by these devices.
1395 	 */
1396 	PCI_CLRBIT(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CMD, PCIM_CMD_MWIEN);
1397 
1398 
1399 #ifdef __brokenalpha__
1400 	/*
1401 	 * Must insure that we do not cross an 8K (bytes) boundary
1402 	 * for DMA reads.  Our highest limit is 1K bytes.  This is a
1403 	 * restriction on some ALPHA platforms with early revision
1404 	 * 21174 PCI chipsets, such as the AlphaPC 164lx
1405 	 */
1406 	PCI_SETBIT(sc, BGE_PCI_DMA_RW_CTL, BGE_PCI_READ_BNDRY_1024, 4);
1407 #endif
1408 
1409 	/* Set the timer prescaler (always 66MHz) */
1410 	CSR_WRITE_4(sc, BGE_MISC_CFG, 65 << 1/*BGE_32BITTIME_66MHZ*/);
1411 
1412 	return(0);
1413 }
1414 
1415 static int
1416 bge_blockinit(struct bge_softc *sc)
1417 {
1418 	volatile struct bge_rcb		*rcb;
1419 	bus_size_t		rcb_addr;
1420 	int			i;
1421 	struct ifnet		*ifp = &sc->ethercom.ec_if;
1422 	bge_hostaddr		taddr;
1423 
1424 	/*
1425 	 * Initialize the memory window pointer register so that
1426 	 * we can access the first 32K of internal NIC RAM. This will
1427 	 * allow us to set up the TX send ring RCBs and the RX return
1428 	 * ring RCBs, plus other things which live in NIC memory.
1429 	 */
1430 
1431 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MEMWIN_BASEADDR, 0);
1432 
1433 	/* Configure mbuf memory pool */
1434 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1435 		if (sc->bge_extram) {
1436 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_BASEADDR,
1437 			    BGE_EXT_SSRAM);
1438 			if ((sc->bge_quirks & BGE_QUIRK_FEWER_MBUFS) != 0)
1439 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x10000);
1440 			else
1441 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x18000);
1442 		} else {
1443 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_BASEADDR,
1444 			    BGE_BUFFPOOL_1);
1445 			if ((sc->bge_quirks & BGE_QUIRK_FEWER_MBUFS) != 0)
1446 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x10000);
1447 			else
1448 				CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_LEN, 0x18000);
1449 		}
1450 
1451 		/* Configure DMA resource pool */
1452 		CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_BASEADDR,
1453 		    BGE_DMA_DESCRIPTORS);
1454 		CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_LEN, 0x2000);
1455 	}
1456 
1457 	/* Configure mbuf pool watermarks */
1458 #ifdef ORIG_WPAUL_VALUES
1459 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 24);
1460 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 24);
1461 	CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 48);
1462 #else
1463 	/* new broadcom docs strongly recommend these: */
1464 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1465 		if (ifp->if_mtu > ETHER_MAX_LEN) {
1466 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 0x50);
1467 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 0x20);
1468 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 0x60);
1469 		} else {
1470 			/* Values from Linux driver... */
1471 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 304);
1472 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 152);
1473 			CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 380);
1474 		}
1475 	} else {
1476 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_READDMA_LOWAT, 0x0);
1477 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_MACRX_LOWAT, 0x10);
1478 		CSR_WRITE_4(sc, BGE_BMAN_MBUFPOOL_HIWAT, 0x60);
1479 	}
1480 #endif
1481 
1482 	/* Configure DMA resource watermarks */
1483 	CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_LOWAT, 5);
1484 	CSR_WRITE_4(sc, BGE_BMAN_DMA_DESCPOOL_HIWAT, 10);
1485 
1486 	/* Enable buffer manager */
1487 	CSR_WRITE_4(sc, BGE_BMAN_MODE,
1488 	    BGE_BMANMODE_ENABLE|BGE_BMANMODE_LOMBUF_ATTN);
1489 
1490 	/* Poll for buffer manager start indication */
1491 	for (i = 0; i < BGE_TIMEOUT; i++) {
1492 		if (CSR_READ_4(sc, BGE_BMAN_MODE) & BGE_BMANMODE_ENABLE)
1493 			break;
1494 		DELAY(10);
1495 	}
1496 
1497 	if (i == BGE_TIMEOUT) {
1498 		aprint_error_dev(sc->bge_dev,
1499 		    "buffer manager failed to start\n");
1500 		return(ENXIO);
1501 	}
1502 
1503 	/* Enable flow-through queues */
1504 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0xFFFFFFFF);
1505 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0);
1506 
1507 	/* Wait until queue initialization is complete */
1508 	for (i = 0; i < BGE_TIMEOUT; i++) {
1509 		if (CSR_READ_4(sc, BGE_FTQ_RESET) == 0)
1510 			break;
1511 		DELAY(10);
1512 	}
1513 
1514 	if (i == BGE_TIMEOUT) {
1515 		aprint_error_dev(sc->bge_dev,
1516 		    "flow-through queue init failed\n");
1517 		return(ENXIO);
1518 	}
1519 
1520 	/* Initialize the standard RX ring control block */
1521 	rcb = &sc->bge_rdata->bge_info.bge_std_rx_rcb;
1522 	bge_set_hostaddr(&rcb->bge_hostaddr,
1523 	    BGE_RING_DMA_ADDR(sc, bge_rx_std_ring));
1524 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1525 		rcb->bge_maxlen_flags =
1526 		    BGE_RCB_MAXLEN_FLAGS(BGE_MAX_FRAMELEN, 0);
1527 	} else {
1528 		rcb->bge_maxlen_flags = BGE_RCB_MAXLEN_FLAGS(512, 0);
1529 	}
1530 	if (sc->bge_extram)
1531 		rcb->bge_nicaddr = BGE_EXT_STD_RX_RINGS;
1532 	else
1533 		rcb->bge_nicaddr = BGE_STD_RX_RINGS;
1534 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_HADDR_HI, rcb->bge_hostaddr.bge_addr_hi);
1535 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_HADDR_LO, rcb->bge_hostaddr.bge_addr_lo);
1536 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_MAXLEN_FLAGS, rcb->bge_maxlen_flags);
1537 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_NICADDR, rcb->bge_nicaddr);
1538 
1539 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1540 		sc->bge_return_ring_cnt = BGE_RETURN_RING_CNT;
1541 	} else {
1542 		sc->bge_return_ring_cnt = BGE_RETURN_RING_CNT_5705;
1543 	}
1544 
1545 	/*
1546 	 * Initialize the jumbo RX ring control block
1547 	 * We set the 'ring disabled' bit in the flags
1548 	 * field until we're actually ready to start
1549 	 * using this ring (i.e. once we set the MTU
1550 	 * high enough to require it).
1551 	 */
1552 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1553 		rcb = &sc->bge_rdata->bge_info.bge_jumbo_rx_rcb;
1554 		bge_set_hostaddr(&rcb->bge_hostaddr,
1555 		    BGE_RING_DMA_ADDR(sc, bge_rx_jumbo_ring));
1556 		rcb->bge_maxlen_flags =
1557 		    BGE_RCB_MAXLEN_FLAGS(BGE_MAX_FRAMELEN,
1558 			BGE_RCB_FLAG_RING_DISABLED);
1559 		if (sc->bge_extram)
1560 			rcb->bge_nicaddr = BGE_EXT_JUMBO_RX_RINGS;
1561 		else
1562 			rcb->bge_nicaddr = BGE_JUMBO_RX_RINGS;
1563 
1564 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_HADDR_HI,
1565 		    rcb->bge_hostaddr.bge_addr_hi);
1566 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_HADDR_LO,
1567 		    rcb->bge_hostaddr.bge_addr_lo);
1568 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_MAXLEN_FLAGS,
1569 		    rcb->bge_maxlen_flags);
1570 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_NICADDR, rcb->bge_nicaddr);
1571 
1572 		/* Set up dummy disabled mini ring RCB */
1573 		rcb = &sc->bge_rdata->bge_info.bge_mini_rx_rcb;
1574 		rcb->bge_maxlen_flags = BGE_RCB_MAXLEN_FLAGS(0,
1575 		    BGE_RCB_FLAG_RING_DISABLED);
1576 		CSR_WRITE_4(sc, BGE_RX_MINI_RCB_MAXLEN_FLAGS,
1577 		    rcb->bge_maxlen_flags);
1578 
1579 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
1580 		    offsetof(struct bge_ring_data, bge_info),
1581 		    sizeof (struct bge_gib),
1582 		    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1583 	}
1584 
1585 	/*
1586 	 * Set the BD ring replenish thresholds. The recommended
1587 	 * values are 1/8th the number of descriptors allocated to
1588 	 * each ring.
1589 	 */
1590 	i = BGE_STD_RX_RING_CNT / 8;
1591 
1592 	/*
1593  	 * Use a value of 8 for the following chips to workaround HW errata.
1594 	 * Some of these chips have been added based on empirical
1595 	 * evidence (they don't work unless this is done).
1596 	 */
1597 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5750 ||
1598 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5752 ||
1599 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5755 ||
1600 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5787)
1601 		i = 8;
1602 
1603 	CSR_WRITE_4(sc, BGE_RBDI_STD_REPL_THRESH, i);
1604 	CSR_WRITE_4(sc, BGE_RBDI_JUMBO_REPL_THRESH, BGE_JUMBO_RX_RING_CNT/8);
1605 
1606 	/*
1607 	 * Disable all unused send rings by setting the 'ring disabled'
1608 	 * bit in the flags field of all the TX send ring control blocks.
1609 	 * These are located in NIC memory.
1610 	 */
1611 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
1612 	for (i = 0; i < BGE_TX_RINGS_EXTSSRAM_MAX; i++) {
1613 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1614 		    BGE_RCB_MAXLEN_FLAGS(0,BGE_RCB_FLAG_RING_DISABLED));
1615 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
1616 		rcb_addr += sizeof(struct bge_rcb);
1617 	}
1618 
1619 	/* Configure TX RCB 0 (we use only the first ring) */
1620 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
1621 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_tx_ring));
1622 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
1623 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
1624 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr,
1625 		    BGE_NIC_TXRING_ADDR(0, BGE_TX_RING_CNT));
1626 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1627 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1628 		    BGE_RCB_MAXLEN_FLAGS(BGE_TX_RING_CNT, 0));
1629 	}
1630 
1631 	/* Disable all unused RX return rings */
1632 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
1633 	for (i = 0; i < BGE_RX_RINGS_MAX; i++) {
1634 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, 0);
1635 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, 0);
1636 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1637 			    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt,
1638                                      BGE_RCB_FLAG_RING_DISABLED));
1639 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
1640 		CSR_WRITE_4(sc, BGE_MBX_RX_CONS0_LO +
1641 		    (i * (sizeof(u_int64_t))), 0);
1642 		rcb_addr += sizeof(struct bge_rcb);
1643 	}
1644 
1645 	/* Initialize RX ring indexes */
1646 	CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, 0);
1647 	CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, 0);
1648 	CSR_WRITE_4(sc, BGE_MBX_RX_MINI_PROD_LO, 0);
1649 
1650 	/*
1651 	 * Set up RX return ring 0
1652 	 * Note that the NIC address for RX return rings is 0x00000000.
1653 	 * The return rings live entirely within the host, so the
1654 	 * nicaddr field in the RCB isn't used.
1655 	 */
1656 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
1657 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_rx_return_ring));
1658 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
1659 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
1660 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0x00000000);
1661 	RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1662 	    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt, 0));
1663 
1664 	/* Set random backoff seed for TX */
1665 	CSR_WRITE_4(sc, BGE_TX_RANDOM_BACKOFF,
1666 	    CLLADDR(ifp->if_sadl)[0] + CLLADDR(ifp->if_sadl)[1] +
1667 	    CLLADDR(ifp->if_sadl)[2] + CLLADDR(ifp->if_sadl)[3] +
1668 	    CLLADDR(ifp->if_sadl)[4] + CLLADDR(ifp->if_sadl)[5] +
1669 	    BGE_TX_BACKOFF_SEED_MASK);
1670 
1671 	/* Set inter-packet gap */
1672 	CSR_WRITE_4(sc, BGE_TX_LENGTHS, 0x2620);
1673 
1674 	/*
1675 	 * Specify which ring to use for packets that don't match
1676 	 * any RX rules.
1677 	 */
1678 	CSR_WRITE_4(sc, BGE_RX_RULES_CFG, 0x08);
1679 
1680 	/*
1681 	 * Configure number of RX lists. One interrupt distribution
1682 	 * list, sixteen active lists, one bad frames class.
1683 	 */
1684 	CSR_WRITE_4(sc, BGE_RXLP_CFG, 0x181);
1685 
1686 	/* Inialize RX list placement stats mask. */
1687 	CSR_WRITE_4(sc, BGE_RXLP_STATS_ENABLE_MASK, 0x007FFFFF);
1688 	CSR_WRITE_4(sc, BGE_RXLP_STATS_CTL, 0x1);
1689 
1690 	/* Disable host coalescing until we get it set up */
1691 	CSR_WRITE_4(sc, BGE_HCC_MODE, 0x00000000);
1692 
1693 	/* Poll to make sure it's shut down. */
1694 	for (i = 0; i < BGE_TIMEOUT; i++) {
1695 		if (!(CSR_READ_4(sc, BGE_HCC_MODE) & BGE_HCCMODE_ENABLE))
1696 			break;
1697 		DELAY(10);
1698 	}
1699 
1700 	if (i == BGE_TIMEOUT) {
1701 		aprint_error_dev(sc->bge_dev,
1702 		    "host coalescing engine failed to idle\n");
1703 		return(ENXIO);
1704 	}
1705 
1706 	/* Set up host coalescing defaults */
1707 	CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, sc->bge_rx_coal_ticks);
1708 	CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS, sc->bge_tx_coal_ticks);
1709 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, sc->bge_rx_max_coal_bds);
1710 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS, sc->bge_tx_max_coal_bds);
1711 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1712 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS_INT, 0);
1713 		CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS_INT, 0);
1714 	}
1715 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS_INT, 0);
1716 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS_INT, 0);
1717 
1718 	/* Set up address of statistics block */
1719 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1720 		bge_set_hostaddr(&taddr,
1721 		    BGE_RING_DMA_ADDR(sc, bge_info.bge_stats));
1722 		CSR_WRITE_4(sc, BGE_HCC_STATS_TICKS, sc->bge_stat_ticks);
1723 		CSR_WRITE_4(sc, BGE_HCC_STATS_BASEADDR, BGE_STATS_BLOCK);
1724 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_HI, taddr.bge_addr_hi);
1725 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_LO, taddr.bge_addr_lo);
1726 	}
1727 
1728 	/* Set up address of status block */
1729 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_status_block));
1730 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_BASEADDR, BGE_STATUS_BLOCK);
1731 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_HI, taddr.bge_addr_hi);
1732 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_LO, taddr.bge_addr_lo);
1733 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx = 0;
1734 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx = 0;
1735 
1736 	/* Turn on host coalescing state machine */
1737 	CSR_WRITE_4(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
1738 
1739 	/* Turn on RX BD completion state machine and enable attentions */
1740 	CSR_WRITE_4(sc, BGE_RBDC_MODE,
1741 	    BGE_RBDCMODE_ENABLE|BGE_RBDCMODE_ATTN);
1742 
1743 	/* Turn on RX list placement state machine */
1744 	CSR_WRITE_4(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
1745 
1746 	/* Turn on RX list selector state machine. */
1747 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1748 		CSR_WRITE_4(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
1749 	}
1750 
1751 	/* Turn on DMA, clear stats */
1752 	CSR_WRITE_4(sc, BGE_MAC_MODE, BGE_MACMODE_TXDMA_ENB|
1753 	    BGE_MACMODE_RXDMA_ENB|BGE_MACMODE_RX_STATS_CLEAR|
1754 	    BGE_MACMODE_TX_STATS_CLEAR|BGE_MACMODE_RX_STATS_ENB|
1755 	    BGE_MACMODE_TX_STATS_ENB|BGE_MACMODE_FRMHDR_DMA_ENB|
1756 	    (sc->bge_tbi ? BGE_PORTMODE_TBI : BGE_PORTMODE_MII));
1757 
1758 	/* Set misc. local control, enable interrupts on attentions */
1759 	sc->bge_local_ctrl_reg = BGE_MLC_INTR_ONATTN | BGE_MLC_AUTO_EEPROM;
1760 
1761 #ifdef notdef
1762 	/* Assert GPIO pins for PHY reset */
1763 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUT0|
1764 	    BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUT2);
1765 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUTEN0|
1766 	    BGE_MLC_MISCIO_OUTEN1|BGE_MLC_MISCIO_OUTEN2);
1767 #endif
1768 
1769 #if defined(not_quite_yet)
1770 	/* Linux driver enables enable gpio pin #1 on 5700s */
1771 	if (sc->bge_chipid == BGE_CHIPID_BCM5700) {
1772 		sc->bge_local_ctrl_reg |=
1773 		  (BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUTEN1);
1774 	}
1775 #endif
1776 	CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
1777 
1778 	/* Turn on DMA completion state machine */
1779 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1780 		CSR_WRITE_4(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
1781 	}
1782 
1783 	/* Turn on write DMA state machine */
1784 	{
1785 		uint32_t bge_wdma_mode =
1786 			BGE_WDMAMODE_ENABLE|BGE_WDMAMODE_ALL_ATTNS;
1787 
1788 		if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5755 ||
1789 		    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5787)
1790 		  /* Enable host coalescing bug fix; see Linux tg3.c */
1791 		  bge_wdma_mode |= (1 << 29);
1792 
1793 		CSR_WRITE_4(sc, BGE_WDMA_MODE, bge_wdma_mode);
1794         }
1795 
1796 	/* Turn on read DMA state machine */
1797 	{
1798 		uint32_t dma_read_modebits;
1799 
1800 		dma_read_modebits =
1801 		  BGE_RDMAMODE_ENABLE | BGE_RDMAMODE_ALL_ATTNS;
1802 
1803 		if (sc->bge_pcie && 0) {
1804 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_LONG_BURST;
1805 		} else if ((sc->bge_quirks & BGE_QUIRK_5705_CORE)) {
1806 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_SIZE_128;
1807 		}
1808 
1809 		/* XXX broadcom-supplied linux driver; undocumented */
1810 		if (BGE_IS_5750_OR_BEYOND(sc)) {
1811  			/*
1812 			 * XXX: magic values.
1813 			 * From Broadcom-supplied Linux driver;  apparently
1814 			 * required to workaround a DMA bug affecting TSO
1815 			 * on bcm575x/bcm5721?
1816 			 */
1817 			dma_read_modebits |= (1 << 27);
1818 		}
1819 		CSR_WRITE_4(sc, BGE_RDMA_MODE, dma_read_modebits);
1820 	}
1821 
1822 	/* Turn on RX data completion state machine */
1823 	CSR_WRITE_4(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
1824 
1825 	/* Turn on RX BD initiator state machine */
1826 	CSR_WRITE_4(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
1827 
1828 	/* Turn on RX data and RX BD initiator state machine */
1829 	CSR_WRITE_4(sc, BGE_RDBDI_MODE, BGE_RDBDIMODE_ENABLE);
1830 
1831 	/* Turn on Mbuf cluster free state machine */
1832 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1833 		CSR_WRITE_4(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
1834 	}
1835 
1836 	/* Turn on send BD completion state machine */
1837 	CSR_WRITE_4(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
1838 
1839 	/* Turn on send data completion state machine */
1840 	CSR_WRITE_4(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
1841 
1842 	/* Turn on send data initiator state machine */
1843 	if (BGE_IS_5750_OR_BEYOND(sc)) {
1844 		/* XXX: magic value from Linux driver */
1845 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE | 0x08);
1846 	} else {
1847 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
1848 	}
1849 
1850 	/* Turn on send BD initiator state machine */
1851 	CSR_WRITE_4(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
1852 
1853 	/* Turn on send BD selector state machine */
1854 	CSR_WRITE_4(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
1855 
1856 	CSR_WRITE_4(sc, BGE_SDI_STATS_ENABLE_MASK, 0x007FFFFF);
1857 	CSR_WRITE_4(sc, BGE_SDI_STATS_CTL,
1858 	    BGE_SDISTATSCTL_ENABLE|BGE_SDISTATSCTL_FASTER);
1859 
1860 	/* ack/clear link change events */
1861 	CSR_WRITE_4(sc, BGE_MAC_STS, BGE_MACSTAT_SYNC_CHANGED|
1862 	    BGE_MACSTAT_CFG_CHANGED);
1863 	CSR_WRITE_4(sc, BGE_MI_STS, 0);
1864 
1865 	/* Enable PHY auto polling (for MII/GMII only) */
1866 	if (sc->bge_tbi) {
1867 		CSR_WRITE_4(sc, BGE_MI_STS, BGE_MISTS_LINK);
1868  	} else {
1869 		BGE_SETBIT(sc, BGE_MI_MODE, BGE_MIMODE_AUTOPOLL|10<<16);
1870 		if (sc->bge_quirks & BGE_QUIRK_LINK_STATE_BROKEN)
1871 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
1872 			    BGE_EVTENB_MI_INTERRUPT);
1873 	}
1874 
1875 	/* Enable link state change attentions. */
1876 	BGE_SETBIT(sc, BGE_MAC_EVT_ENB, BGE_EVTENB_LINK_CHANGED);
1877 
1878 	return(0);
1879 }
1880 
1881 static const struct bge_revision {
1882 	uint32_t		br_chipid;
1883 	uint32_t		br_quirks;
1884 	const char		*br_name;
1885 } bge_revisions[] = {
1886 	{ BGE_CHIPID_BCM5700_A0,
1887 	  BGE_QUIRK_LINK_STATE_BROKEN,
1888 	  "BCM5700 A0" },
1889 
1890 	{ BGE_CHIPID_BCM5700_A1,
1891 	  BGE_QUIRK_LINK_STATE_BROKEN,
1892 	  "BCM5700 A1" },
1893 
1894 	{ BGE_CHIPID_BCM5700_B0,
1895 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_CSUM_BROKEN|BGE_QUIRK_5700_COMMON,
1896 	  "BCM5700 B0" },
1897 
1898 	{ BGE_CHIPID_BCM5700_B1,
1899 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
1900 	  "BCM5700 B1" },
1901 
1902 	{ BGE_CHIPID_BCM5700_B2,
1903 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
1904 	  "BCM5700 B2" },
1905 
1906 	{ BGE_CHIPID_BCM5700_B3,
1907 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
1908 	  "BCM5700 B3" },
1909 
1910 	/* This is treated like a BCM5700 Bx */
1911 	{ BGE_CHIPID_BCM5700_ALTIMA,
1912 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
1913 	  "BCM5700 Altima" },
1914 
1915 	{ BGE_CHIPID_BCM5700_C0,
1916 	  0,
1917 	  "BCM5700 C0" },
1918 
1919 	{ BGE_CHIPID_BCM5701_A0,
1920 	  0, /*XXX really, just not known */
1921 	  "BCM5701 A0" },
1922 
1923 	{ BGE_CHIPID_BCM5701_B0,
1924 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
1925 	  "BCM5701 B0" },
1926 
1927 	{ BGE_CHIPID_BCM5701_B2,
1928 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
1929 	  "BCM5701 B2" },
1930 
1931 	{ BGE_CHIPID_BCM5701_B5,
1932 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
1933 	  "BCM5701 B5" },
1934 
1935 	{ BGE_CHIPID_BCM5703_A0,
1936 	  0,
1937 	  "BCM5703 A0" },
1938 
1939 	{ BGE_CHIPID_BCM5703_A1,
1940 	  0,
1941 	  "BCM5703 A1" },
1942 
1943 	{ BGE_CHIPID_BCM5703_A2,
1944 	  BGE_QUIRK_ONLY_PHY_1,
1945 	  "BCM5703 A2" },
1946 
1947 	{ BGE_CHIPID_BCM5703_A3,
1948 	  BGE_QUIRK_ONLY_PHY_1,
1949 	  "BCM5703 A3" },
1950 
1951 	{ BGE_CHIPID_BCM5703_B0,
1952 	  BGE_QUIRK_ONLY_PHY_1,
1953 	  "BCM5703 B0" },
1954 
1955 	{ BGE_CHIPID_BCM5704_A0,
1956   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1957 	  "BCM5704 A0" },
1958 
1959 	{ BGE_CHIPID_BCM5704_A1,
1960   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1961 	  "BCM5704 A1" },
1962 
1963 	{ BGE_CHIPID_BCM5704_A2,
1964   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1965 	  "BCM5704 A2" },
1966 
1967 	{ BGE_CHIPID_BCM5704_A3,
1968   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1969 	  "BCM5704 A3" },
1970 
1971 	{ BGE_CHIPID_BCM5705_A0,
1972 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1973 	  "BCM5705 A0" },
1974 
1975 	{ BGE_CHIPID_BCM5705_A1,
1976 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1977 	  "BCM5705 A1" },
1978 
1979 	{ BGE_CHIPID_BCM5705_A2,
1980 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1981 	  "BCM5705 A2" },
1982 
1983 	{ BGE_CHIPID_BCM5705_A3,
1984 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1985 	  "BCM5705 A3" },
1986 
1987 	{ BGE_CHIPID_BCM5750_A0,
1988 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1989 	  "BCM5750 A0" },
1990 
1991 	{ BGE_CHIPID_BCM5750_A1,
1992 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1993 	  "BCM5750 A1" },
1994 
1995 	{ BGE_CHIPID_BCM5751_A1,
1996 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1997 	  "BCM5751 A1" },
1998 
1999 	{ BGE_CHIPID_BCM5752_A0,
2000 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2001 	  "BCM5752 A0" },
2002 
2003 	{ BGE_CHIPID_BCM5752_A1,
2004 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2005 	  "BCM5752 A1" },
2006 
2007 	{ BGE_CHIPID_BCM5752_A2,
2008 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2009 	  "BCM5752 A2" },
2010 
2011 	{ BGE_CHIPID_BCM5755_A0,
2012 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2013 	  "BCM5755 A0" },
2014 
2015 	{ BGE_CHIPID_BCM5755_A1,
2016 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2017 	  "BCM5755 A1" },
2018 
2019 	{ BGE_CHIPID_BCM5755_A2,
2020 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2021 	  "BCM5755 A2" },
2022 
2023 	{ BGE_CHIPID_BCM5755_C0,
2024 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2025 	  "BCM5755 C0" },
2026 
2027 	{ BGE_CHIPID_BCM5787_A0,
2028 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2029 	  "BCM5754/5787 A0" },
2030 
2031 	{ BGE_CHIPID_BCM5787_A1,
2032 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2033 	  "BCM5754/5787 A1" },
2034 
2035 	{ BGE_CHIPID_BCM5787_A2,
2036 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2037 	  "BCM5754/5787 A2" },
2038 
2039 	{ 0, 0, NULL }
2040 };
2041 
2042 /*
2043  * Some defaults for major revisions, so that newer steppings
2044  * that we don't know about have a shot at working.
2045  */
2046 static const struct bge_revision bge_majorrevs[] = {
2047 	{ BGE_ASICREV_BCM5700,
2048 	  BGE_QUIRK_LINK_STATE_BROKEN,
2049 	  "unknown BCM5700" },
2050 
2051 	{ BGE_ASICREV_BCM5701,
2052 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
2053 	  "unknown BCM5701" },
2054 
2055 	{ BGE_ASICREV_BCM5703,
2056 	  0,
2057 	  "unknown BCM5703" },
2058 
2059 	{ BGE_ASICREV_BCM5704,
2060 	  BGE_QUIRK_ONLY_PHY_1,
2061 	  "unknown BCM5704" },
2062 
2063 	{ BGE_ASICREV_BCM5705,
2064 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2065 	  "unknown BCM5705" },
2066 
2067 	{ BGE_ASICREV_BCM5750,
2068 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2069 	  "unknown BCM575x family" },
2070 
2071 	{ BGE_ASICREV_BCM5714_A0,
2072 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2073 	  "unknown BCM5714" },
2074 
2075 	{ BGE_ASICREV_BCM5714,
2076 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2077 	  "unknown BCM5714" },
2078 
2079 	{ BGE_ASICREV_BCM5752,
2080 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2081 	  "unknown BCM5752 family" },
2082 
2083 	{ BGE_ASICREV_BCM5755,
2084 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2085 	  "unknown BCM5755" },
2086 
2087 	{ BGE_ASICREV_BCM5780,
2088 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2089 	  "unknown BCM5780" },
2090 
2091 	{ BGE_ASICREV_BCM5787,
2092 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2093 	  "unknown BCM5787" },
2094 
2095 	{ 0,
2096 	  0,
2097 	  NULL }
2098 };
2099 
2100 
2101 static const struct bge_revision *
2102 bge_lookup_rev(uint32_t chipid)
2103 {
2104 	const struct bge_revision *br;
2105 
2106 	for (br = bge_revisions; br->br_name != NULL; br++) {
2107 		if (br->br_chipid == chipid)
2108 			return (br);
2109 	}
2110 
2111 	for (br = bge_majorrevs; br->br_name != NULL; br++) {
2112 		if (br->br_chipid == BGE_ASICREV(chipid))
2113 			return (br);
2114 	}
2115 
2116 	return (NULL);
2117 }
2118 
2119 static const struct bge_product {
2120 	pci_vendor_id_t		bp_vendor;
2121 	pci_product_id_t	bp_product;
2122 	const char		*bp_name;
2123 } bge_products[] = {
2124 	/*
2125 	 * The BCM5700 documentation seems to indicate that the hardware
2126 	 * still has the Alteon vendor ID burned into it, though it
2127 	 * should always be overridden by the value in the EEPROM.  We'll
2128 	 * check for it anyway.
2129 	 */
2130 	{ PCI_VENDOR_ALTEON,
2131 	  PCI_PRODUCT_ALTEON_BCM5700,
2132 	  "Broadcom BCM5700 Gigabit Ethernet",
2133 	  },
2134 	{ PCI_VENDOR_ALTEON,
2135 	  PCI_PRODUCT_ALTEON_BCM5701,
2136 	  "Broadcom BCM5701 Gigabit Ethernet",
2137 	  },
2138 
2139 	{ PCI_VENDOR_ALTIMA,
2140 	  PCI_PRODUCT_ALTIMA_AC1000,
2141 	  "Altima AC1000 Gigabit Ethernet",
2142 	  },
2143 	{ PCI_VENDOR_ALTIMA,
2144 	  PCI_PRODUCT_ALTIMA_AC1001,
2145 	  "Altima AC1001 Gigabit Ethernet",
2146 	   },
2147 	{ PCI_VENDOR_ALTIMA,
2148 	  PCI_PRODUCT_ALTIMA_AC9100,
2149 	  "Altima AC9100 Gigabit Ethernet",
2150 	  },
2151 
2152 	{ PCI_VENDOR_BROADCOM,
2153 	  PCI_PRODUCT_BROADCOM_BCM5700,
2154 	  "Broadcom BCM5700 Gigabit Ethernet",
2155 	  },
2156 	{ PCI_VENDOR_BROADCOM,
2157 	  PCI_PRODUCT_BROADCOM_BCM5701,
2158 	  "Broadcom BCM5701 Gigabit Ethernet",
2159 	  },
2160 	{ PCI_VENDOR_BROADCOM,
2161 	  PCI_PRODUCT_BROADCOM_BCM5702,
2162 	  "Broadcom BCM5702 Gigabit Ethernet",
2163 	  },
2164 	{ PCI_VENDOR_BROADCOM,
2165 	  PCI_PRODUCT_BROADCOM_BCM5702X,
2166 	  "Broadcom BCM5702X Gigabit Ethernet" },
2167 
2168 	{ PCI_VENDOR_BROADCOM,
2169 	  PCI_PRODUCT_BROADCOM_BCM5703,
2170 	  "Broadcom BCM5703 Gigabit Ethernet",
2171 	  },
2172 	{ PCI_VENDOR_BROADCOM,
2173 	  PCI_PRODUCT_BROADCOM_BCM5703X,
2174 	  "Broadcom BCM5703X Gigabit Ethernet",
2175 	  },
2176 	{ PCI_VENDOR_BROADCOM,
2177 	  PCI_PRODUCT_BROADCOM_BCM5703_ALT,
2178 	  "Broadcom BCM5703 Gigabit Ethernet",
2179 	  },
2180 
2181    	{ PCI_VENDOR_BROADCOM,
2182 	  PCI_PRODUCT_BROADCOM_BCM5704C,
2183 	  "Broadcom BCM5704C Dual Gigabit Ethernet",
2184 	  },
2185    	{ PCI_VENDOR_BROADCOM,
2186 	  PCI_PRODUCT_BROADCOM_BCM5704S,
2187 	  "Broadcom BCM5704S Dual Gigabit Ethernet",
2188 	  },
2189 
2190    	{ PCI_VENDOR_BROADCOM,
2191 	  PCI_PRODUCT_BROADCOM_BCM5705,
2192 	  "Broadcom BCM5705 Gigabit Ethernet",
2193 	  },
2194    	{ PCI_VENDOR_BROADCOM,
2195 	  PCI_PRODUCT_BROADCOM_BCM5705K,
2196 	  "Broadcom BCM5705K Gigabit Ethernet",
2197 	  },
2198    	{ PCI_VENDOR_BROADCOM,
2199 	  PCI_PRODUCT_BROADCOM_BCM5705M,
2200 	  "Broadcom BCM5705M Gigabit Ethernet",
2201 	  },
2202    	{ PCI_VENDOR_BROADCOM,
2203 	  PCI_PRODUCT_BROADCOM_BCM5705M_ALT,
2204 	  "Broadcom BCM5705M Gigabit Ethernet",
2205 	  },
2206 
2207 	{ PCI_VENDOR_BROADCOM,
2208 	  PCI_PRODUCT_BROADCOM_BCM5714,
2209 	  "Broadcom BCM5714/5715 Gigabit Ethernet",
2210 	  },
2211 	{ PCI_VENDOR_BROADCOM,
2212 	  PCI_PRODUCT_BROADCOM_BCM5715,
2213 	  "Broadcom BCM5714/5715 Gigabit Ethernet",
2214 	  },
2215 	{ PCI_VENDOR_BROADCOM,
2216 	  PCI_PRODUCT_BROADCOM_BCM5789,
2217 	  "Broadcom BCM5789 Gigabit Ethernet",
2218 	  },
2219 
2220 	{ PCI_VENDOR_BROADCOM,
2221 	  PCI_PRODUCT_BROADCOM_BCM5721,
2222 	  "Broadcom BCM5721 Gigabit Ethernet",
2223 	  },
2224 
2225 	{ PCI_VENDOR_BROADCOM,
2226 	  PCI_PRODUCT_BROADCOM_BCM5722,
2227 	  "Broadcom BCM5722 Gigabit Ethernet",
2228 	  },
2229 
2230 	{ PCI_VENDOR_BROADCOM,
2231 	  PCI_PRODUCT_BROADCOM_BCM5750,
2232 	  "Broadcom BCM5750 Gigabit Ethernet",
2233 	  },
2234 
2235 	{ PCI_VENDOR_BROADCOM,
2236 	  PCI_PRODUCT_BROADCOM_BCM5750M,
2237 	  "Broadcom BCM5750M Gigabit Ethernet",
2238 	  },
2239 
2240 	{ PCI_VENDOR_BROADCOM,
2241 	  PCI_PRODUCT_BROADCOM_BCM5751,
2242 	  "Broadcom BCM5751 Gigabit Ethernet",
2243 	  },
2244 
2245 	{ PCI_VENDOR_BROADCOM,
2246 	  PCI_PRODUCT_BROADCOM_BCM5751M,
2247 	  "Broadcom BCM5751M Gigabit Ethernet",
2248 	  },
2249 
2250 	{ PCI_VENDOR_BROADCOM,
2251 	  PCI_PRODUCT_BROADCOM_BCM5752,
2252 	  "Broadcom BCM5752 Gigabit Ethernet",
2253 	  },
2254 
2255 	{ PCI_VENDOR_BROADCOM,
2256 	  PCI_PRODUCT_BROADCOM_BCM5752M,
2257 	  "Broadcom BCM5752M Gigabit Ethernet",
2258 	  },
2259 
2260 	{ PCI_VENDOR_BROADCOM,
2261 	  PCI_PRODUCT_BROADCOM_BCM5753,
2262 	  "Broadcom BCM5753 Gigabit Ethernet",
2263 	  },
2264 
2265 	{ PCI_VENDOR_BROADCOM,
2266 	  PCI_PRODUCT_BROADCOM_BCM5753M,
2267 	  "Broadcom BCM5753M Gigabit Ethernet",
2268 	  },
2269 
2270 	{ PCI_VENDOR_BROADCOM,
2271 	  PCI_PRODUCT_BROADCOM_BCM5754,
2272 	  "Broadcom BCM5754 Gigabit Ethernet",
2273 	},
2274 
2275 	{ PCI_VENDOR_BROADCOM,
2276 	  PCI_PRODUCT_BROADCOM_BCM5754M,
2277 	  "Broadcom BCM5754M Gigabit Ethernet",
2278 	},
2279 
2280 	{ PCI_VENDOR_BROADCOM,
2281 	  PCI_PRODUCT_BROADCOM_BCM5755,
2282 	  "Broadcom BCM5755 Gigabit Ethernet",
2283 	},
2284 
2285 	{ PCI_VENDOR_BROADCOM,
2286 	  PCI_PRODUCT_BROADCOM_BCM5755M,
2287 	  "Broadcom BCM5755M Gigabit Ethernet",
2288 	},
2289 
2290    	{ PCI_VENDOR_BROADCOM,
2291 	  PCI_PRODUCT_BROADCOM_BCM5780,
2292 	  "Broadcom BCM5780 Gigabit Ethernet",
2293 	  },
2294 
2295    	{ PCI_VENDOR_BROADCOM,
2296 	  PCI_PRODUCT_BROADCOM_BCM5780S,
2297 	  "Broadcom BCM5780S Gigabit Ethernet",
2298 	  },
2299 
2300    	{ PCI_VENDOR_BROADCOM,
2301 	  PCI_PRODUCT_BROADCOM_BCM5782,
2302 	  "Broadcom BCM5782 Gigabit Ethernet",
2303 	},
2304 
2305 	{ PCI_VENDOR_BROADCOM,
2306 	  PCI_PRODUCT_BROADCOM_BCM5786,
2307 	  "Broadcom BCM5786 Gigabit Ethernet",
2308 	},
2309 
2310 	{ PCI_VENDOR_BROADCOM,
2311 	  PCI_PRODUCT_BROADCOM_BCM5787,
2312 	  "Broadcom BCM5787 Gigabit Ethernet",
2313 	},
2314 
2315 	{ PCI_VENDOR_BROADCOM,
2316 	  PCI_PRODUCT_BROADCOM_BCM5787M,
2317 	  "Broadcom BCM5787M Gigabit Ethernet",
2318 	},
2319 
2320    	{ PCI_VENDOR_BROADCOM,
2321 	  PCI_PRODUCT_BROADCOM_BCM5788,
2322 	  "Broadcom BCM5788 Gigabit Ethernet",
2323 	  },
2324    	{ PCI_VENDOR_BROADCOM,
2325 	  PCI_PRODUCT_BROADCOM_BCM5789,
2326 	  "Broadcom BCM5789 Gigabit Ethernet",
2327 	  },
2328 
2329    	{ PCI_VENDOR_BROADCOM,
2330 	  PCI_PRODUCT_BROADCOM_BCM5901,
2331 	  "Broadcom BCM5901 Fast Ethernet",
2332 	  },
2333    	{ PCI_VENDOR_BROADCOM,
2334 	  PCI_PRODUCT_BROADCOM_BCM5901A2,
2335 	  "Broadcom BCM5901A2 Fast Ethernet",
2336 	  },
2337 
2338 	{ PCI_VENDOR_SCHNEIDERKOCH,
2339 	  PCI_PRODUCT_SCHNEIDERKOCH_SK_9DX1,
2340 	  "SysKonnect SK-9Dx1 Gigabit Ethernet",
2341 	  },
2342 
2343 	{ PCI_VENDOR_3COM,
2344 	  PCI_PRODUCT_3COM_3C996,
2345 	  "3Com 3c996 Gigabit Ethernet",
2346 	  },
2347 
2348 	{ 0,
2349 	  0,
2350 	  NULL },
2351 };
2352 
2353 static const struct bge_product *
2354 bge_lookup(const struct pci_attach_args *pa)
2355 {
2356 	const struct bge_product *bp;
2357 
2358 	for (bp = bge_products; bp->bp_name != NULL; bp++) {
2359 		if (PCI_VENDOR(pa->pa_id) == bp->bp_vendor &&
2360 		    PCI_PRODUCT(pa->pa_id) == bp->bp_product)
2361 			return (bp);
2362 	}
2363 
2364 	return (NULL);
2365 }
2366 
2367 static int
2368 bge_setpowerstate(struct bge_softc *sc, int powerlevel)
2369 {
2370 #ifdef NOTYET
2371 	u_int32_t pm_ctl = 0;
2372 
2373 	/* XXX FIXME: make sure indirect accesses enabled? */
2374 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_MISC_CTL, 4);
2375 	pm_ctl |= BGE_PCIMISCCTL_INDIRECT_ACCESS;
2376 	pci_write_config(sc->bge_dev, BGE_PCI_MISC_CTL, pm_ctl, 4);
2377 
2378 	/* clear the PME_assert bit and power state bits, enable PME */
2379 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_PWRMGMT_CMD, 2);
2380 	pm_ctl &= ~PCIM_PSTAT_DMASK;
2381 	pm_ctl |= (1 << 8);
2382 
2383 	if (powerlevel == 0) {
2384 		pm_ctl |= PCIM_PSTAT_D0;
2385 		pci_write_config(sc->bge_dev, BGE_PCI_PWRMGMT_CMD,
2386 		    pm_ctl, 2);
2387 		DELAY(10000);
2388 		CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
2389 		DELAY(10000);
2390 
2391 #ifdef NOTYET
2392 		/* XXX FIXME: write 0x02 to phy aux_Ctrl reg */
2393 		bge_miibus_writereg(sc->bge_dev, 1, 0x18, 0x02);
2394 #endif
2395 		DELAY(40); DELAY(40); DELAY(40);
2396 		DELAY(10000);	/* above not quite adequate on 5700 */
2397 		return 0;
2398 	}
2399 
2400 
2401 	/*
2402 	 * Entering ACPI power states D1-D3 is achieved by wiggling
2403 	 * GMII gpio pins. Example code assumes all hardware vendors
2404 	 * followed Broadom's sample pcb layout. Until we verify that
2405 	 * for all supported OEM cards, states D1-D3 are  unsupported.
2406 	 */
2407 	aprint_error_dev(sc->bge_dev,
2408 	    "power state %d unimplemented; check GPIO pins\n",
2409 	    powerlevel);
2410 #endif
2411 	return EOPNOTSUPP;
2412 }
2413 
2414 
2415 /*
2416  * Probe for a Broadcom chip. Check the PCI vendor and device IDs
2417  * against our list and return its name if we find a match. Note
2418  * that since the Broadcom controller contains VPD support, we
2419  * can get the device name string from the controller itself instead
2420  * of the compiled-in string. This is a little slow, but it guarantees
2421  * we'll always announce the right product name.
2422  */
2423 static int
2424 bge_probe(device_t parent, cfdata_t match, void *aux)
2425 {
2426 	struct pci_attach_args *pa = (struct pci_attach_args *)aux;
2427 
2428 	if (bge_lookup(pa) != NULL)
2429 		return (1);
2430 
2431 	return (0);
2432 }
2433 
2434 static void
2435 bge_attach(device_t parent, device_t self, void *aux)
2436 {
2437 	struct bge_softc	*sc = device_private(self);
2438 	struct pci_attach_args	*pa = aux;
2439 	const struct bge_product *bp;
2440 	const struct bge_revision *br;
2441 	pci_chipset_tag_t	pc;
2442 	pci_intr_handle_t	ih;
2443 	const char		*intrstr = NULL;
2444 	bus_dma_segment_t	seg;
2445 	int			rseg;
2446 	u_int32_t		hwcfg = 0;
2447 	u_int32_t		mac_addr = 0;
2448 	u_int32_t		command;
2449 	struct ifnet		*ifp;
2450 	void *			kva;
2451 	u_char			eaddr[ETHER_ADDR_LEN];
2452 	pcireg_t		memtype;
2453 	bus_addr_t		memaddr;
2454 	bus_size_t		memsize;
2455 	u_int32_t		pm_ctl;
2456 
2457 	bp = bge_lookup(pa);
2458 	KASSERT(bp != NULL);
2459 
2460 	sc->sc_pc = pa->pa_pc;
2461 	sc->sc_pcitag = pa->pa_tag;
2462 	sc->bge_dev = self;
2463 
2464 	aprint_naive(": Ethernet controller\n");
2465 	aprint_normal(": %s\n", bp->bp_name);
2466 
2467 	/*
2468 	 * Map control/status registers.
2469 	 */
2470 	DPRINTFN(5, ("Map control/status regs\n"));
2471 	pc = sc->sc_pc;
2472 	command = pci_conf_read(pc, sc->sc_pcitag, PCI_COMMAND_STATUS_REG);
2473 	command |= PCI_COMMAND_MEM_ENABLE | PCI_COMMAND_MASTER_ENABLE;
2474 	pci_conf_write(pc, sc->sc_pcitag, PCI_COMMAND_STATUS_REG, command);
2475 	command = pci_conf_read(pc, sc->sc_pcitag, PCI_COMMAND_STATUS_REG);
2476 
2477 	if (!(command & PCI_COMMAND_MEM_ENABLE)) {
2478 		aprint_error_dev(sc->bge_dev,
2479 		    "failed to enable memory mapping!\n");
2480 		return;
2481 	}
2482 
2483 	DPRINTFN(5, ("pci_mem_find\n"));
2484 	memtype = pci_mapreg_type(sc->sc_pc, sc->sc_pcitag, BGE_PCI_BAR0);
2485  	switch (memtype) {
2486 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_32BIT:
2487 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_64BIT:
2488 		if (pci_mapreg_map(pa, BGE_PCI_BAR0,
2489 		    memtype, 0, &sc->bge_btag, &sc->bge_bhandle,
2490 		    &memaddr, &memsize) == 0)
2491 			break;
2492 	default:
2493 		aprint_error_dev(sc->bge_dev, "can't find mem space\n");
2494 		return;
2495 	}
2496 
2497 	DPRINTFN(5, ("pci_intr_map\n"));
2498 	if (pci_intr_map(pa, &ih)) {
2499 		aprint_error_dev(sc->bge_dev, "couldn't map interrupt\n");
2500 		return;
2501 	}
2502 
2503 	DPRINTFN(5, ("pci_intr_string\n"));
2504 	intrstr = pci_intr_string(pc, ih);
2505 
2506 	DPRINTFN(5, ("pci_intr_establish\n"));
2507 	sc->bge_intrhand = pci_intr_establish(pc, ih, IPL_NET, bge_intr, sc);
2508 
2509 	if (sc->bge_intrhand == NULL) {
2510 		aprint_error_dev(sc->bge_dev,
2511 		    "couldn't establish interrupt%s%s\n",
2512 		    intrstr ? " at " : "", intrstr ? intrstr : "");
2513 		return;
2514 	}
2515 	aprint_normal_dev(sc->bge_dev, "interrupting at %s\n", intrstr);
2516 
2517 	/*
2518 	 * Kludge for 5700 Bx bug: a hardware bug (PCIX byte enable?)
2519 	 * can clobber the chip's PCI config-space power control registers,
2520 	 * leaving the card in D3 powersave state.
2521 	 * We do not have memory-mapped registers in this state,
2522 	 * so force device into D0 state before starting initialization.
2523 	 */
2524 	pm_ctl = pci_conf_read(pc, sc->sc_pcitag, BGE_PCI_PWRMGMT_CMD);
2525 	pm_ctl &= ~(PCI_PWR_D0|PCI_PWR_D1|PCI_PWR_D2|PCI_PWR_D3);
2526 	pm_ctl |= (1 << 8) | PCI_PWR_D0 ; /* D0 state */
2527 	pci_conf_write(pc, sc->sc_pcitag, BGE_PCI_PWRMGMT_CMD, pm_ctl);
2528 	DELAY(1000);	/* 27 usec is allegedly sufficent */
2529 
2530 	/*
2531 	 * Save ASIC rev.  Look up any quirks associated with this
2532 	 * ASIC.
2533 	 */
2534 	sc->bge_chipid =
2535 	    pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MISC_CTL) &
2536 	    BGE_PCIMISCCTL_ASICREV;
2537 
2538 	/*
2539 	 * Detect PCI-Express devices
2540 	 * XXX: guessed from Linux/FreeBSD; no documentation
2541 	 */
2542 	if (pci_get_capability(sc->sc_pc, sc->sc_pcitag, PCI_CAP_PCIEXPRESS,
2543 	        NULL, NULL) != 0)
2544 		sc->bge_pcie = 1;
2545 	else
2546 		sc->bge_pcie = 0;
2547 
2548 	/* Try to reset the chip. */
2549 	DPRINTFN(5, ("bge_reset\n"));
2550 	bge_reset(sc);
2551 
2552 	if (bge_chipinit(sc)) {
2553 		aprint_error_dev(sc->bge_dev, "chip initialization failed\n");
2554 		bge_release_resources(sc);
2555 		return;
2556 	}
2557 
2558 	/*
2559 	 * Get station address from the EEPROM.
2560 	 */
2561 	mac_addr = bge_readmem_ind(sc, 0x0c14);
2562 	if ((mac_addr >> 16) == 0x484b) {
2563 		eaddr[0] = (u_char)(mac_addr >> 8);
2564 		eaddr[1] = (u_char)(mac_addr >> 0);
2565 		mac_addr = bge_readmem_ind(sc, 0x0c18);
2566 		eaddr[2] = (u_char)(mac_addr >> 24);
2567 		eaddr[3] = (u_char)(mac_addr >> 16);
2568 		eaddr[4] = (u_char)(mac_addr >> 8);
2569 		eaddr[5] = (u_char)(mac_addr >> 0);
2570 	} else if (bge_read_eeprom(sc, (void *)eaddr,
2571 	    BGE_EE_MAC_OFFSET + 2, ETHER_ADDR_LEN)) {
2572 		aprint_error_dev(sc->bge_dev,
2573 		    "failed to read station address\n");
2574 		bge_release_resources(sc);
2575 		return;
2576 	}
2577 
2578 	br = bge_lookup_rev(sc->bge_chipid);
2579 
2580 	if (br == NULL) {
2581 		aprint_normal_dev(sc->bge_dev, "unknown ASIC (0x%04x)",
2582 		    sc->bge_chipid >> 16);
2583 		sc->bge_quirks = 0;
2584 	} else {
2585 		aprint_normal_dev(sc->bge_dev, "ASIC %s (0x%04x)",
2586 		    br->br_name, sc->bge_chipid >> 16);
2587 		sc->bge_quirks |= br->br_quirks;
2588 	}
2589 	aprint_normal(", Ethernet address %s\n", ether_sprintf(eaddr));
2590 
2591 	/* Allocate the general information block and ring buffers. */
2592 	if (pci_dma64_available(pa))
2593 		sc->bge_dmatag = pa->pa_dmat64;
2594 	else
2595 		sc->bge_dmatag = pa->pa_dmat;
2596 	DPRINTFN(5, ("bus_dmamem_alloc\n"));
2597 	if (bus_dmamem_alloc(sc->bge_dmatag, sizeof(struct bge_ring_data),
2598 			     PAGE_SIZE, 0, &seg, 1, &rseg, BUS_DMA_NOWAIT)) {
2599 		aprint_error_dev(sc->bge_dev, "can't alloc rx buffers\n");
2600 		return;
2601 	}
2602 	DPRINTFN(5, ("bus_dmamem_map\n"));
2603 	if (bus_dmamem_map(sc->bge_dmatag, &seg, rseg,
2604 			   sizeof(struct bge_ring_data), &kva,
2605 			   BUS_DMA_NOWAIT)) {
2606 		aprint_error_dev(sc->bge_dev,
2607 		    "can't map DMA buffers (%zu bytes)\n",
2608 		    sizeof(struct bge_ring_data));
2609 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2610 		return;
2611 	}
2612 	DPRINTFN(5, ("bus_dmamem_create\n"));
2613 	if (bus_dmamap_create(sc->bge_dmatag, sizeof(struct bge_ring_data), 1,
2614 	    sizeof(struct bge_ring_data), 0,
2615 	    BUS_DMA_NOWAIT, &sc->bge_ring_map)) {
2616 		aprint_error_dev(sc->bge_dev, "can't create DMA map\n");
2617 		bus_dmamem_unmap(sc->bge_dmatag, kva,
2618 				 sizeof(struct bge_ring_data));
2619 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2620 		return;
2621 	}
2622 	DPRINTFN(5, ("bus_dmamem_load\n"));
2623 	if (bus_dmamap_load(sc->bge_dmatag, sc->bge_ring_map, kva,
2624 			    sizeof(struct bge_ring_data), NULL,
2625 			    BUS_DMA_NOWAIT)) {
2626 		bus_dmamap_destroy(sc->bge_dmatag, sc->bge_ring_map);
2627 		bus_dmamem_unmap(sc->bge_dmatag, kva,
2628 				 sizeof(struct bge_ring_data));
2629 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2630 		return;
2631 	}
2632 
2633 	DPRINTFN(5, ("bzero\n"));
2634 	sc->bge_rdata = (struct bge_ring_data *)kva;
2635 
2636 	memset(sc->bge_rdata, 0, sizeof(struct bge_ring_data));
2637 
2638 	/* Try to allocate memory for jumbo buffers. */
2639 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
2640 		if (bge_alloc_jumbo_mem(sc)) {
2641 			aprint_error_dev(sc->bge_dev,
2642 			    "jumbo buffer allocation failed\n");
2643 		} else
2644 			sc->ethercom.ec_capabilities |= ETHERCAP_JUMBO_MTU;
2645 	}
2646 
2647 	/* Set default tuneable values. */
2648 	sc->bge_stat_ticks = BGE_TICKS_PER_SEC;
2649 	sc->bge_rx_coal_ticks = 150;
2650 	sc->bge_rx_max_coal_bds = 64;
2651 #ifdef ORIG_WPAUL_VALUES
2652 	sc->bge_tx_coal_ticks = 150;
2653 	sc->bge_tx_max_coal_bds = 128;
2654 #else
2655 	sc->bge_tx_coal_ticks = 300;
2656 	sc->bge_tx_max_coal_bds = 400;
2657 #endif
2658 	if (sc->bge_quirks & BGE_QUIRK_5705_CORE) {
2659 		sc->bge_tx_coal_ticks = (12 * 5);
2660 		sc->bge_tx_max_coal_bds = (12 * 5);
2661 			aprint_verbose_dev(sc->bge_dev,
2662 			    "setting short Tx thresholds\n");
2663 	}
2664 
2665 	/* Set up ifnet structure */
2666 	ifp = &sc->ethercom.ec_if;
2667 	ifp->if_softc = sc;
2668 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
2669 	ifp->if_ioctl = bge_ioctl;
2670 	ifp->if_stop = bge_stop;
2671 	ifp->if_start = bge_start;
2672 	ifp->if_init = bge_init;
2673 	ifp->if_watchdog = bge_watchdog;
2674 	IFQ_SET_MAXLEN(&ifp->if_snd, max(BGE_TX_RING_CNT - 1, IFQ_MAXLEN));
2675 	IFQ_SET_READY(&ifp->if_snd);
2676 	DPRINTFN(5, ("strcpy if_xname\n"));
2677 	strcpy(ifp->if_xname, device_xname(sc->bge_dev));
2678 
2679 	if ((sc->bge_quirks & BGE_QUIRK_CSUM_BROKEN) == 0)
2680 		sc->ethercom.ec_if.if_capabilities |=
2681 		    IFCAP_CSUM_IPv4_Tx | IFCAP_CSUM_IPv4_Rx |
2682 		    IFCAP_CSUM_TCPv4_Tx | IFCAP_CSUM_TCPv4_Rx |
2683 		    IFCAP_CSUM_UDPv4_Tx | IFCAP_CSUM_UDPv4_Rx;
2684 	sc->ethercom.ec_capabilities |=
2685 	    ETHERCAP_VLAN_HWTAGGING | ETHERCAP_VLAN_MTU;
2686 
2687 	if (sc->bge_pcie)
2688 		sc->ethercom.ec_if.if_capabilities |= IFCAP_TSOv4;
2689 
2690 	/*
2691 	 * Do MII setup.
2692 	 */
2693 	DPRINTFN(5, ("mii setup\n"));
2694 	sc->bge_mii.mii_ifp = ifp;
2695 	sc->bge_mii.mii_readreg = bge_miibus_readreg;
2696 	sc->bge_mii.mii_writereg = bge_miibus_writereg;
2697 	sc->bge_mii.mii_statchg = bge_miibus_statchg;
2698 
2699 	/*
2700 	 * Figure out what sort of media we have by checking the
2701 	 * hardware config word in the first 32k of NIC internal memory,
2702 	 * or fall back to the config word in the EEPROM. Note: on some BCM5700
2703 	 * cards, this value appears to be unset. If that's the
2704 	 * case, we have to rely on identifying the NIC by its PCI
2705 	 * subsystem ID, as we do below for the SysKonnect SK-9D41.
2706 	 */
2707 	if (bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_SIG) == BGE_MAGIC_NUMBER) {
2708 		hwcfg = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_NICCFG);
2709 	} else {
2710 		bge_read_eeprom(sc, (void *)&hwcfg,
2711 		    BGE_EE_HWCFG_OFFSET, sizeof(hwcfg));
2712 		hwcfg = be32toh(hwcfg);
2713 	}
2714 	if ((hwcfg & BGE_HWCFG_MEDIA) == BGE_MEDIA_FIBER)
2715 		sc->bge_tbi = 1;
2716 
2717 	/* The SysKonnect SK-9D41 is a 1000baseSX card. */
2718 	if ((pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_SUBSYS) >> 16) ==
2719 	    SK_SUBSYSID_9D41)
2720 		sc->bge_tbi = 1;
2721 
2722 	if (sc->bge_tbi) {
2723 		ifmedia_init(&sc->bge_ifmedia, IFM_IMASK, bge_ifmedia_upd,
2724 		    bge_ifmedia_sts);
2725 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX, 0, NULL);
2726 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX|IFM_FDX,
2727 			    0, NULL);
2728 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO, 0, NULL);
2729 		ifmedia_set(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO);
2730 	} else {
2731 		/*
2732 		 * Do transceiver setup.
2733 		 */
2734 		ifmedia_init(&sc->bge_mii.mii_media, 0, bge_ifmedia_upd,
2735 			     bge_ifmedia_sts);
2736 		mii_attach(sc->bge_dev, &sc->bge_mii, 0xffffffff,
2737 			   MII_PHY_ANY, MII_OFFSET_ANY,
2738 			   MIIF_FORCEANEG|MIIF_DOPAUSE);
2739 
2740 		if (LIST_EMPTY(&sc->bge_mii.mii_phys)) {
2741 			aprint_error_dev(sc->bge_dev, "no PHY found!\n");
2742 			ifmedia_add(&sc->bge_mii.mii_media,
2743 				    IFM_ETHER|IFM_MANUAL, 0, NULL);
2744 			ifmedia_set(&sc->bge_mii.mii_media,
2745 				    IFM_ETHER|IFM_MANUAL);
2746 		} else
2747 			ifmedia_set(&sc->bge_mii.mii_media,
2748 				    IFM_ETHER|IFM_AUTO);
2749 	}
2750 
2751 	/*
2752 	 * When using the BCM5701 in PCI-X mode, data corruption has
2753 	 * been observed in the first few bytes of some received packets.
2754 	 * Aligning the packet buffer in memory eliminates the corruption.
2755 	 * Unfortunately, this misaligns the packet payloads.  On platforms
2756 	 * which do not support unaligned accesses, we will realign the
2757 	 * payloads by copying the received packets.
2758 	 */
2759 	if (sc->bge_quirks & BGE_QUIRK_PCIX_DMA_ALIGN_BUG) {
2760 		/* If in PCI-X mode, work around the alignment bug. */
2761 		if ((pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_PCISTATE) &
2762                     (BGE_PCISTATE_PCI_BUSMODE | BGE_PCISTATE_PCI_BUSSPEED)) ==
2763                          BGE_PCISTATE_PCI_BUSSPEED)
2764 		sc->bge_rx_alignment_bug = 1;
2765         }
2766 
2767 	/*
2768 	 * Call MI attach routine.
2769 	 */
2770 	DPRINTFN(5, ("if_attach\n"));
2771 	if_attach(ifp);
2772 	DPRINTFN(5, ("ether_ifattach\n"));
2773 	ether_ifattach(ifp, eaddr);
2774 #if NRND > 0
2775 	rnd_attach_source(&sc->rnd_source, device_xname(sc->bge_dev),
2776 		RND_TYPE_NET, 0);
2777 #endif
2778 #ifdef BGE_EVENT_COUNTERS
2779 	/*
2780 	 * Attach event counters.
2781 	 */
2782 	evcnt_attach_dynamic(&sc->bge_ev_intr, EVCNT_TYPE_INTR,
2783 	    NULL, device_xname(sc->bge_dev), "intr");
2784 	evcnt_attach_dynamic(&sc->bge_ev_tx_xoff, EVCNT_TYPE_MISC,
2785 	    NULL, device_xname(sc->bge_dev), "tx_xoff");
2786 	evcnt_attach_dynamic(&sc->bge_ev_tx_xon, EVCNT_TYPE_MISC,
2787 	    NULL, device_xname(sc->bge_dev), "tx_xon");
2788 	evcnt_attach_dynamic(&sc->bge_ev_rx_xoff, EVCNT_TYPE_MISC,
2789 	    NULL, device_xname(sc->bge_dev), "rx_xoff");
2790 	evcnt_attach_dynamic(&sc->bge_ev_rx_xon, EVCNT_TYPE_MISC,
2791 	    NULL, device_xname(sc->bge_dev), "rx_xon");
2792 	evcnt_attach_dynamic(&sc->bge_ev_rx_macctl, EVCNT_TYPE_MISC,
2793 	    NULL, device_xname(sc->bge_dev), "rx_macctl");
2794 	evcnt_attach_dynamic(&sc->bge_ev_xoffentered, EVCNT_TYPE_MISC,
2795 	    NULL, device_xname(sc->bge_dev), "xoffentered");
2796 #endif /* BGE_EVENT_COUNTERS */
2797 	DPRINTFN(5, ("callout_init\n"));
2798 	callout_init(&sc->bge_timeout, 0);
2799 
2800 	if (!pmf_device_register(self, NULL, NULL))
2801 		aprint_error_dev(self, "couldn't establish power handler\n");
2802 	else
2803 		pmf_class_network_register(self, ifp);
2804 }
2805 
2806 static void
2807 bge_release_resources(struct bge_softc *sc)
2808 {
2809 	if (sc->bge_vpd_prodname != NULL)
2810 		free(sc->bge_vpd_prodname, M_DEVBUF);
2811 
2812 	if (sc->bge_vpd_readonly != NULL)
2813 		free(sc->bge_vpd_readonly, M_DEVBUF);
2814 }
2815 
2816 static void
2817 bge_reset(struct bge_softc *sc)
2818 {
2819 	u_int32_t cachesize, command, pcistate, new_pcistate;
2820 	int i, val;
2821 
2822 	/* Save some important PCI state. */
2823 	cachesize = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CACHESZ);
2824 	command = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CMD);
2825 	pcistate = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_PCISTATE);
2826 
2827 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MISC_CTL,
2828 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
2829 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
2830 
2831 	/*
2832 	 * Disable the firmware fastboot feature on 5752 ASIC
2833 	 * to avoid firmware timeout.
2834 	 */
2835 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5752 ||
2836 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5755 ||
2837 	    BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5787)
2838 		CSR_WRITE_4(sc, BGE_FASTBOOT_PC, 0);
2839 
2840 	val = BGE_MISCCFG_RESET_CORE_CLOCKS | (65<<1);
2841 	/*
2842 	 * XXX: from FreeBSD/Linux; no documentation
2843 	 */
2844 	if (sc->bge_pcie) {
2845 		if (CSR_READ_4(sc, BGE_PCIE_CTL1) == 0x60)
2846 			CSR_WRITE_4(sc, BGE_PCIE_CTL1, 0x20);
2847 		if (sc->bge_chipid != BGE_CHIPID_BCM5750_A0) {
2848 			/* No idea what that actually means */
2849 			CSR_WRITE_4(sc, BGE_MISC_CFG, 1 << 29);
2850 			val |= (1<<29);
2851 		}
2852 	}
2853 
2854 	/* Issue global reset */
2855 	bge_writereg_ind(sc, BGE_MISC_CFG, val);
2856 
2857 	DELAY(1000);
2858 
2859 	/*
2860 	 * XXX: from FreeBSD/Linux; no documentation
2861 	 */
2862 	if (sc->bge_pcie) {
2863 		if (sc->bge_chipid == BGE_CHIPID_BCM5750_A0) {
2864 			pcireg_t reg;
2865 
2866 			DELAY(500000);
2867 			/* XXX: Magic Numbers */
2868 			reg = pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_UNKNOWN0);
2869 			pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_UNKNOWN0,
2870 			    reg | (1 << 15));
2871 		}
2872 		/*
2873 		 * XXX: Magic Numbers.
2874 		 * Sets maximal PCI-e payload and clears any PCI-e errors.
2875 		 * Should be replaced with references to PCI config-space
2876 		 * capability block for PCI-Express.
2877 		 */
2878 		pci_conf_write(sc->sc_pc, sc->sc_pcitag,
2879 		    BGE_PCI_CONF_DEV_CTRL, 0xf5000);
2880 
2881 	}
2882 
2883 	/* Reset some of the PCI state that got zapped by reset */
2884 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_MISC_CTL,
2885 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
2886 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
2887 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CMD, command);
2888 	pci_conf_write(sc->sc_pc, sc->sc_pcitag, BGE_PCI_CACHESZ, cachesize);
2889 	bge_writereg_ind(sc, BGE_MISC_CFG, (65 << 1));
2890 
2891 	/* Enable memory arbiter. */
2892 	{
2893 		uint32_t marbmode = 0;
2894 		if (BGE_IS_5714_FAMILY(sc)) {
2895 			marbmode = CSR_READ_4(sc, BGE_MARB_MODE);
2896 		}
2897  		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE | marbmode);
2898 	}
2899 
2900 	/*
2901 	 * Write the magic number to the firmware mailbox at 0xb50
2902 	 * so that the driver can synchronize with the firmware.
2903 	 */
2904 	bge_writemem_ind(sc, BGE_SOFTWARE_GENCOMM, BGE_MAGIC_NUMBER);
2905 
2906 	/*
2907 	 * Poll the value location we just wrote until
2908 	 * we see the 1's complement of the magic number.
2909 	 * This indicates that the firmware initialization
2910 	 * is complete.
2911 	 */
2912 	for (i = 0; i < BGE_TIMEOUT; i++) {
2913 		val = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM);
2914 		if (val == ~BGE_MAGIC_NUMBER)
2915 			break;
2916 		DELAY(1000);
2917 	}
2918 
2919 	if (i >= BGE_TIMEOUT) {
2920 		aprint_error_dev(sc->bge_dev,
2921 		    "firmware handshake timed out, val = %x\n", val);
2922 		/*
2923 		 * XXX: occasionally fired on bcm5721, but without
2924 		 * apparent harm.  For now, keep going if we timeout
2925 		 * against PCI-E devices.
2926 		 */
2927 		 if (!sc->bge_pcie)
2928 		  return;
2929 	}
2930 
2931 	/*
2932 	 * XXX Wait for the value of the PCISTATE register to
2933 	 * return to its original pre-reset state. This is a
2934 	 * fairly good indicator of reset completion. If we don't
2935 	 * wait for the reset to fully complete, trying to read
2936 	 * from the device's non-PCI registers may yield garbage
2937 	 * results.
2938 	 */
2939 	for (i = 0; i < 10000; i++) {
2940 		new_pcistate = pci_conf_read(sc->sc_pc, sc->sc_pcitag,
2941 		    BGE_PCI_PCISTATE);
2942 		if ((new_pcistate & ~BGE_PCISTATE_RESERVED) ==
2943 		    (pcistate & ~BGE_PCISTATE_RESERVED))
2944 			break;
2945 		DELAY(10);
2946 	}
2947 	if ((new_pcistate & ~BGE_PCISTATE_RESERVED) !=
2948 	    (pcistate & ~BGE_PCISTATE_RESERVED)) {
2949 		aprint_error_dev(sc->bge_dev, "pcistate failed to revert\n");
2950 	}
2951 
2952 	/* XXX: from FreeBSD/Linux; no documentation */
2953 	if (sc->bge_pcie && sc->bge_chipid != BGE_CHIPID_BCM5750_A0)
2954 		CSR_WRITE_4(sc, BGE_PCIE_CTL0, CSR_READ_4(sc, BGE_PCIE_CTL0) | (1<<25));
2955 
2956 	/* Enable memory arbiter. */
2957 	/* XXX why do this twice? */
2958 	{
2959 		uint32_t marbmode = 0;
2960 		if (BGE_IS_5714_FAMILY(sc)) {
2961 			marbmode = CSR_READ_4(sc, BGE_MARB_MODE);
2962 		}
2963  		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE | marbmode);
2964 	}
2965 
2966 	/* Fix up byte swapping */
2967 	CSR_WRITE_4(sc, BGE_MODE_CTL, BGE_DMA_SWAP_OPTIONS);
2968 
2969 	CSR_WRITE_4(sc, BGE_MAC_MODE, 0);
2970 
2971 	DELAY(10000);
2972 }
2973 
2974 /*
2975  * Frame reception handling. This is called if there's a frame
2976  * on the receive return list.
2977  *
2978  * Note: we have to be able to handle two possibilities here:
2979  * 1) the frame is from the jumbo recieve ring
2980  * 2) the frame is from the standard receive ring
2981  */
2982 
2983 static void
2984 bge_rxeof(struct bge_softc *sc)
2985 {
2986 	struct ifnet *ifp;
2987 	int stdcnt = 0, jumbocnt = 0;
2988 	bus_dmamap_t dmamap;
2989 	bus_addr_t offset, toff;
2990 	bus_size_t tlen;
2991 	int tosync;
2992 
2993 	ifp = &sc->ethercom.ec_if;
2994 
2995 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2996 	    offsetof(struct bge_ring_data, bge_status_block),
2997 	    sizeof (struct bge_status_block),
2998 	    BUS_DMASYNC_POSTREAD);
2999 
3000 	offset = offsetof(struct bge_ring_data, bge_rx_return_ring);
3001 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx -
3002 	    sc->bge_rx_saved_considx;
3003 
3004 #if NRND > 0
3005 	if (tosync != 0 && RND_ENABLED(&sc->rnd_source))
3006 		rnd_add_uint32(&sc->rnd_source, tosync);
3007 #endif
3008 
3009 	toff = offset + (sc->bge_rx_saved_considx * sizeof (struct bge_rx_bd));
3010 
3011 	if (tosync < 0) {
3012 		tlen = (sc->bge_return_ring_cnt - sc->bge_rx_saved_considx) *
3013 		    sizeof (struct bge_rx_bd);
3014 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
3015 		    toff, tlen, BUS_DMASYNC_POSTREAD);
3016 		tosync = -tosync;
3017 	}
3018 
3019 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
3020 	    offset, tosync * sizeof (struct bge_rx_bd),
3021 	    BUS_DMASYNC_POSTREAD);
3022 
3023 	while(sc->bge_rx_saved_considx !=
3024 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx) {
3025 		struct bge_rx_bd	*cur_rx;
3026 		u_int32_t		rxidx;
3027 		struct mbuf		*m = NULL;
3028 
3029 		cur_rx = &sc->bge_rdata->
3030 			bge_rx_return_ring[sc->bge_rx_saved_considx];
3031 
3032 		rxidx = cur_rx->bge_idx;
3033 		BGE_INC(sc->bge_rx_saved_considx, sc->bge_return_ring_cnt);
3034 
3035 		if (cur_rx->bge_flags & BGE_RXBDFLAG_JUMBO_RING) {
3036 			BGE_INC(sc->bge_jumbo, BGE_JUMBO_RX_RING_CNT);
3037 			m = sc->bge_cdata.bge_rx_jumbo_chain[rxidx];
3038 			sc->bge_cdata.bge_rx_jumbo_chain[rxidx] = NULL;
3039 			jumbocnt++;
3040 			bus_dmamap_sync(sc->bge_dmatag,
3041 			    sc->bge_cdata.bge_rx_jumbo_map,
3042 			    mtod(m, char *) - (char *)sc->bge_cdata.bge_jumbo_buf,
3043 			    BGE_JLEN, BUS_DMASYNC_POSTREAD);
3044 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
3045 				ifp->if_ierrors++;
3046 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
3047 				continue;
3048 			}
3049 			if (bge_newbuf_jumbo(sc, sc->bge_jumbo,
3050 					     NULL)== ENOBUFS) {
3051 				ifp->if_ierrors++;
3052 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
3053 				continue;
3054 			}
3055 		} else {
3056 			BGE_INC(sc->bge_std, BGE_STD_RX_RING_CNT);
3057 			m = sc->bge_cdata.bge_rx_std_chain[rxidx];
3058 
3059 			sc->bge_cdata.bge_rx_std_chain[rxidx] = NULL;
3060 			stdcnt++;
3061 			dmamap = sc->bge_cdata.bge_rx_std_map[rxidx];
3062 			sc->bge_cdata.bge_rx_std_map[rxidx] = 0;
3063 			bus_dmamap_sync(sc->bge_dmatag, dmamap, 0,
3064 			    dmamap->dm_mapsize, BUS_DMASYNC_POSTREAD);
3065 			bus_dmamap_unload(sc->bge_dmatag, dmamap);
3066 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
3067 				ifp->if_ierrors++;
3068 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
3069 				continue;
3070 			}
3071 			if (bge_newbuf_std(sc, sc->bge_std,
3072 			    NULL, dmamap) == ENOBUFS) {
3073 				ifp->if_ierrors++;
3074 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
3075 				continue;
3076 			}
3077 		}
3078 
3079 		ifp->if_ipackets++;
3080 #ifndef __NO_STRICT_ALIGNMENT
3081                 /*
3082                  * XXX: if the 5701 PCIX-Rx-DMA workaround is in effect,
3083                  * the Rx buffer has the layer-2 header unaligned.
3084                  * If our CPU requires alignment, re-align by copying.
3085                  */
3086 		if (sc->bge_rx_alignment_bug) {
3087 			memmove(mtod(m, char *) + ETHER_ALIGN, m->m_data,
3088                                 cur_rx->bge_len);
3089 			m->m_data += ETHER_ALIGN;
3090 		}
3091 #endif
3092 
3093 		m->m_pkthdr.len = m->m_len = cur_rx->bge_len - ETHER_CRC_LEN;
3094 		m->m_pkthdr.rcvif = ifp;
3095 
3096 #if NBPFILTER > 0
3097 		/*
3098 		 * Handle BPF listeners. Let the BPF user see the packet.
3099 		 */
3100 		if (ifp->if_bpf)
3101 			bpf_mtap(ifp->if_bpf, m);
3102 #endif
3103 
3104 		m->m_pkthdr.csum_flags = M_CSUM_IPv4;
3105 
3106 		if ((cur_rx->bge_ip_csum ^ 0xffff) != 0)
3107 			m->m_pkthdr.csum_flags |= M_CSUM_IPv4_BAD;
3108 		/*
3109 		 * Rx transport checksum-offload may also
3110 		 * have bugs with packets which, when transmitted,
3111 		 * were `runts' requiring padding.
3112 		 */
3113 		if (cur_rx->bge_flags & BGE_RXBDFLAG_TCP_UDP_CSUM &&
3114 		    (/* (sc->_bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||*/
3115 		     m->m_pkthdr.len >= ETHER_MIN_NOPAD)) {
3116 			m->m_pkthdr.csum_data =
3117 			    cur_rx->bge_tcp_udp_csum;
3118 			m->m_pkthdr.csum_flags |=
3119 			    (M_CSUM_TCPv4|M_CSUM_UDPv4|
3120 			     M_CSUM_DATA|M_CSUM_NO_PSEUDOHDR);
3121 		}
3122 
3123 		/*
3124 		 * If we received a packet with a vlan tag, pass it
3125 		 * to vlan_input() instead of ether_input().
3126 		 */
3127 		if (cur_rx->bge_flags & BGE_RXBDFLAG_VLAN_TAG) {
3128 			VLAN_INPUT_TAG(ifp, m, cur_rx->bge_vlan_tag, continue);
3129 		}
3130 
3131 		(*ifp->if_input)(ifp, m);
3132 	}
3133 
3134 	CSR_WRITE_4(sc, BGE_MBX_RX_CONS0_LO, sc->bge_rx_saved_considx);
3135 	if (stdcnt)
3136 		CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, sc->bge_std);
3137 	if (jumbocnt)
3138 		CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, sc->bge_jumbo);
3139 }
3140 
3141 static void
3142 bge_txeof(struct bge_softc *sc)
3143 {
3144 	struct bge_tx_bd *cur_tx = NULL;
3145 	struct ifnet *ifp;
3146 	struct txdmamap_pool_entry *dma;
3147 	bus_addr_t offset, toff;
3148 	bus_size_t tlen;
3149 	int tosync;
3150 	struct mbuf *m;
3151 
3152 	ifp = &sc->ethercom.ec_if;
3153 
3154 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
3155 	    offsetof(struct bge_ring_data, bge_status_block),
3156 	    sizeof (struct bge_status_block),
3157 	    BUS_DMASYNC_POSTREAD);
3158 
3159 	offset = offsetof(struct bge_ring_data, bge_tx_ring);
3160 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx -
3161 	    sc->bge_tx_saved_considx;
3162 
3163 #if NRND > 0
3164 	if (tosync != 0 && RND_ENABLED(&sc->rnd_source))
3165 		rnd_add_uint32(&sc->rnd_source, tosync);
3166 #endif
3167 
3168 	toff = offset + (sc->bge_tx_saved_considx * sizeof (struct bge_tx_bd));
3169 
3170 	if (tosync < 0) {
3171 		tlen = (BGE_TX_RING_CNT - sc->bge_tx_saved_considx) *
3172 		    sizeof (struct bge_tx_bd);
3173 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
3174 		    toff, tlen, BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
3175 		tosync = -tosync;
3176 	}
3177 
3178 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
3179 	    offset, tosync * sizeof (struct bge_tx_bd),
3180 	    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
3181 
3182 	/*
3183 	 * Go through our tx ring and free mbufs for those
3184 	 * frames that have been sent.
3185 	 */
3186 	while (sc->bge_tx_saved_considx !=
3187 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx) {
3188 		u_int32_t		idx = 0;
3189 
3190 		idx = sc->bge_tx_saved_considx;
3191 		cur_tx = &sc->bge_rdata->bge_tx_ring[idx];
3192 		if (cur_tx->bge_flags & BGE_TXBDFLAG_END)
3193 			ifp->if_opackets++;
3194 		m = sc->bge_cdata.bge_tx_chain[idx];
3195 		if (m != NULL) {
3196 			sc->bge_cdata.bge_tx_chain[idx] = NULL;
3197 			dma = sc->txdma[idx];
3198 			bus_dmamap_sync(sc->bge_dmatag, dma->dmamap, 0,
3199 			    dma->dmamap->dm_mapsize, BUS_DMASYNC_POSTWRITE);
3200 			bus_dmamap_unload(sc->bge_dmatag, dma->dmamap);
3201 			SLIST_INSERT_HEAD(&sc->txdma_list, dma, link);
3202 			sc->txdma[idx] = NULL;
3203 
3204 			m_freem(m);
3205 		}
3206 		sc->bge_txcnt--;
3207 		BGE_INC(sc->bge_tx_saved_considx, BGE_TX_RING_CNT);
3208 		ifp->if_timer = 0;
3209 	}
3210 
3211 	if (cur_tx != NULL)
3212 		ifp->if_flags &= ~IFF_OACTIVE;
3213 }
3214 
3215 static int
3216 bge_intr(void *xsc)
3217 {
3218 	struct bge_softc *sc;
3219 	struct ifnet *ifp;
3220 
3221 	sc = xsc;
3222 	ifp = &sc->ethercom.ec_if;
3223 
3224 	/*
3225 	 * Ascertain whether the interrupt is from this bge device.
3226 	 * Do the cheap test first.
3227 	 */
3228 	if ((sc->bge_rdata->bge_status_block.bge_status &
3229 	    BGE_STATFLAG_UPDATED) == 0) {
3230 		/*
3231 		 * Sometimes, the interrupt comes in before the
3232 		 * DMA update of the status block (performed prior
3233 		 * to the  interrupt itself) has completed.
3234 		 * In that case, do the (extremely expensive!)
3235 		 * PCI-config-space register read.
3236 		 */
3237 		uint32_t pcistate =
3238 		    pci_conf_read(sc->sc_pc, sc->sc_pcitag, BGE_PCI_PCISTATE);
3239 
3240 		if (pcistate & BGE_PCISTATE_INTR_STATE)
3241 			return (0);
3242 
3243 	}
3244 	/*
3245 	 *  If we reach here, then the interrupt is for us.
3246 	 */
3247 
3248 	/* Ack interrupt and stop others from occuring. */
3249 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 1);
3250 
3251 	BGE_EVCNT_INCR(sc->bge_ev_intr);
3252 
3253 	/*
3254 	 * Process link state changes.
3255 	 * Grrr. The link status word in the status block does
3256 	 * not work correctly on the BCM5700 rev AX and BX chips,
3257 	 * according to all available information. Hence, we have
3258 	 * to enable MII interrupts in order to properly obtain
3259 	 * async link changes. Unfortunately, this also means that
3260 	 * we have to read the MAC status register to detect link
3261 	 * changes, thereby adding an additional register access to
3262 	 * the interrupt handler.
3263 	 */
3264 
3265 	if (sc->bge_quirks & BGE_QUIRK_LINK_STATE_BROKEN) {
3266 		u_int32_t		status;
3267 
3268 		status = CSR_READ_4(sc, BGE_MAC_STS);
3269 		if (status & BGE_MACSTAT_MI_INTERRUPT) {
3270 			sc->bge_link = 0;
3271 			callout_stop(&sc->bge_timeout);
3272 			bge_tick(sc);
3273 			/* Clear the interrupt */
3274 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
3275 			    BGE_EVTENB_MI_INTERRUPT);
3276 			bge_miibus_readreg(sc->bge_dev, 1, BRGPHY_MII_ISR);
3277 			bge_miibus_writereg(sc->bge_dev, 1, BRGPHY_MII_IMR,
3278 			    BRGPHY_INTRS);
3279 		}
3280 	} else {
3281 		u_int32_t		status;
3282 
3283 		status = CSR_READ_4(sc, BGE_MAC_STS);
3284 		if (status & BGE_MACSTAT_LINK_CHANGED) {
3285 			sc->bge_link = 0;
3286 			callout_stop(&sc->bge_timeout);
3287 			bge_tick(sc);
3288 			/* Clear the interrupt */
3289 			CSR_WRITE_4(sc, BGE_MAC_STS, BGE_MACSTAT_SYNC_CHANGED|
3290 			    BGE_MACSTAT_CFG_CHANGED|BGE_MACSTAT_MI_COMPLETE|
3291 			    BGE_MACSTAT_LINK_CHANGED);
3292 		}
3293 	}
3294 
3295 	if (ifp->if_flags & IFF_RUNNING) {
3296 		/* Check RX return ring producer/consumer */
3297 		bge_rxeof(sc);
3298 
3299 		/* Check TX ring producer/consumer */
3300 		bge_txeof(sc);
3301 	}
3302 
3303 	if (sc->bge_pending_rxintr_change) {
3304 		uint32_t rx_ticks = sc->bge_rx_coal_ticks;
3305 		uint32_t rx_bds = sc->bge_rx_max_coal_bds;
3306 		uint32_t junk;
3307 
3308 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, rx_ticks);
3309 		DELAY(10);
3310 		junk = CSR_READ_4(sc, BGE_HCC_RX_COAL_TICKS);
3311 
3312 		CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, rx_bds);
3313 		DELAY(10);
3314 		junk = CSR_READ_4(sc, BGE_HCC_RX_MAX_COAL_BDS);
3315 
3316 		sc->bge_pending_rxintr_change = 0;
3317 	}
3318 	bge_handle_events(sc);
3319 
3320 	/* Re-enable interrupts. */
3321 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 0);
3322 
3323 	if (ifp->if_flags & IFF_RUNNING && !IFQ_IS_EMPTY(&ifp->if_snd))
3324 		bge_start(ifp);
3325 
3326 	return (1);
3327 }
3328 
3329 static void
3330 bge_tick(void *xsc)
3331 {
3332 	struct bge_softc *sc = xsc;
3333 	struct mii_data *mii = &sc->bge_mii;
3334 	int s;
3335 
3336 	s = splnet();
3337 
3338 	bge_stats_update(sc);
3339 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
3340 
3341 	if (sc->bge_tbi) {
3342 		if (CSR_READ_4(sc, BGE_MAC_STS) &
3343 		    BGE_MACSTAT_TBI_PCS_SYNCHED) {
3344 			CSR_WRITE_4(sc, BGE_MAC_STS, 0xFFFFFFFF);
3345 		}
3346 	} else
3347 		mii_tick(mii);
3348 
3349 	splx(s);
3350 }
3351 
3352 static void
3353 bge_stats_update(struct bge_softc *sc)
3354 {
3355 	struct ifnet *ifp = &sc->ethercom.ec_if;
3356 	bus_size_t stats = BGE_MEMWIN_START + BGE_STATS_BLOCK;
3357 	bus_size_t rstats = BGE_RX_STATS;
3358 
3359 #define READ_RSTAT(sc, stats, stat) \
3360 	  CSR_READ_4(sc, stats + offsetof(struct bge_mac_stats_regs, stat))
3361 
3362 	if (sc->bge_quirks & BGE_QUIRK_5705_CORE) {
3363 		ifp->if_collisions +=
3364 		    READ_RSTAT(sc, rstats, dot3StatsSingleCollisionFrames) +
3365 		    READ_RSTAT(sc, rstats, dot3StatsMultipleCollisionFrames) +
3366 		    READ_RSTAT(sc, rstats, dot3StatsExcessiveCollisions) +
3367 		    READ_RSTAT(sc, rstats, dot3StatsLateCollisions);
3368 
3369 		BGE_EVCNT_ADD(sc->bge_ev_tx_xoff,
3370 			      READ_RSTAT(sc, rstats, outXoffSent));
3371 		BGE_EVCNT_ADD(sc->bge_ev_tx_xon,
3372 			      READ_RSTAT(sc, rstats, outXonSent));
3373 		BGE_EVCNT_ADD(sc->bge_ev_rx_xoff,
3374 			      READ_RSTAT(sc, rstats, xoffPauseFramesReceived));
3375 		BGE_EVCNT_ADD(sc->bge_ev_rx_xon,
3376 			      READ_RSTAT(sc, rstats, xonPauseFramesReceived));
3377 		BGE_EVCNT_ADD(sc->bge_ev_rx_macctl,
3378 			      READ_RSTAT(sc, rstats, macControlFramesReceived));
3379 		BGE_EVCNT_ADD(sc->bge_ev_xoffentered,
3380 			      READ_RSTAT(sc, rstats, xoffStateEntered));
3381 		return;
3382 	}
3383 
3384 #undef READ_RSTAT
3385 #define READ_STAT(sc, stats, stat) \
3386 	  CSR_READ_4(sc, stats + offsetof(struct bge_stats, stat))
3387 
3388 	ifp->if_collisions +=
3389 	  (READ_STAT(sc, stats, dot3StatsSingleCollisionFrames.bge_addr_lo) +
3390 	   READ_STAT(sc, stats, dot3StatsMultipleCollisionFrames.bge_addr_lo) +
3391 	   READ_STAT(sc, stats, dot3StatsExcessiveCollisions.bge_addr_lo) +
3392 	   READ_STAT(sc, stats, dot3StatsLateCollisions.bge_addr_lo)) -
3393 	  ifp->if_collisions;
3394 
3395 	BGE_EVCNT_UPD(sc->bge_ev_tx_xoff,
3396 		      READ_STAT(sc, stats, outXoffSent.bge_addr_lo));
3397 	BGE_EVCNT_UPD(sc->bge_ev_tx_xon,
3398 		      READ_STAT(sc, stats, outXonSent.bge_addr_lo));
3399 	BGE_EVCNT_UPD(sc->bge_ev_rx_xoff,
3400 		      READ_STAT(sc, stats,
3401 		      		xoffPauseFramesReceived.bge_addr_lo));
3402 	BGE_EVCNT_UPD(sc->bge_ev_rx_xon,
3403 		      READ_STAT(sc, stats, xonPauseFramesReceived.bge_addr_lo));
3404 	BGE_EVCNT_UPD(sc->bge_ev_rx_macctl,
3405 		      READ_STAT(sc, stats,
3406 		      		macControlFramesReceived.bge_addr_lo));
3407 	BGE_EVCNT_UPD(sc->bge_ev_xoffentered,
3408 		      READ_STAT(sc, stats, xoffStateEntered.bge_addr_lo));
3409 
3410 #undef READ_STAT
3411 
3412 #ifdef notdef
3413 	ifp->if_collisions +=
3414 	   (sc->bge_rdata->bge_info.bge_stats.dot3StatsSingleCollisionFrames +
3415 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsMultipleCollisionFrames +
3416 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsExcessiveCollisions +
3417 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsLateCollisions) -
3418 	   ifp->if_collisions;
3419 #endif
3420 }
3421 
3422 /*
3423  * Pad outbound frame to ETHER_MIN_NOPAD for an unusual reason.
3424  * The bge hardware will pad out Tx runts to ETHER_MIN_NOPAD,
3425  * but when such padded frames employ the  bge IP/TCP checksum offload,
3426  * the hardware checksum assist gives incorrect results (possibly
3427  * from incorporating its own padding into the UDP/TCP checksum; who knows).
3428  * If we pad such runts with zeros, the onboard checksum comes out correct.
3429  */
3430 static inline int
3431 bge_cksum_pad(struct mbuf *pkt)
3432 {
3433 	struct mbuf *last = NULL;
3434 	int padlen;
3435 
3436 	padlen = ETHER_MIN_NOPAD - pkt->m_pkthdr.len;
3437 
3438 	/* if there's only the packet-header and we can pad there, use it. */
3439 	if (pkt->m_pkthdr.len == pkt->m_len &&
3440 	    M_TRAILINGSPACE(pkt) >= padlen) {
3441 		last = pkt;
3442 	} else {
3443 		/*
3444 		 * Walk packet chain to find last mbuf. We will either
3445 		 * pad there, or append a new mbuf and pad it
3446 		 * (thus perhaps avoiding the bcm5700 dma-min bug).
3447 		 */
3448 		for (last = pkt; last->m_next != NULL; last = last->m_next) {
3449 	      	       continue; /* do nothing */
3450 		}
3451 
3452 		/* `last' now points to last in chain. */
3453 		if (M_TRAILINGSPACE(last) < padlen) {
3454 			/* Allocate new empty mbuf, pad it. Compact later. */
3455 			struct mbuf *n;
3456 			MGET(n, M_DONTWAIT, MT_DATA);
3457 			if (n == NULL)
3458 				return ENOBUFS;
3459 			n->m_len = 0;
3460 			last->m_next = n;
3461 			last = n;
3462 		}
3463 	}
3464 
3465 	KDASSERT(!M_READONLY(last));
3466 	KDASSERT(M_TRAILINGSPACE(last) >= padlen);
3467 
3468 	/* Now zero the pad area, to avoid the bge cksum-assist bug */
3469 	memset(mtod(last, char *) + last->m_len, 0, padlen);
3470 	last->m_len += padlen;
3471 	pkt->m_pkthdr.len += padlen;
3472 	return 0;
3473 }
3474 
3475 /*
3476  * Compact outbound packets to avoid bug with DMA segments less than 8 bytes.
3477  */
3478 static inline int
3479 bge_compact_dma_runt(struct mbuf *pkt)
3480 {
3481 	struct mbuf	*m, *prev;
3482 	int 		totlen, prevlen;
3483 
3484 	prev = NULL;
3485 	totlen = 0;
3486 	prevlen = -1;
3487 
3488 	for (m = pkt; m != NULL; prev = m,m = m->m_next) {
3489 		int mlen = m->m_len;
3490 		int shortfall = 8 - mlen ;
3491 
3492 		totlen += mlen;
3493 		if (mlen == 0) {
3494 			continue;
3495 		}
3496 		if (mlen >= 8)
3497 			continue;
3498 
3499 		/* If we get here, mbuf data is too small for DMA engine.
3500 		 * Try to fix by shuffling data to prev or next in chain.
3501 		 * If that fails, do a compacting deep-copy of the whole chain.
3502 		 */
3503 
3504 		/* Internal frag. If fits in prev, copy it there. */
3505 		if (prev && M_TRAILINGSPACE(prev) >= m->m_len) {
3506 		  	memcpy(prev->m_data + prev->m_len, m->m_data, mlen);
3507 			prev->m_len += mlen;
3508 			m->m_len = 0;
3509 			/* XXX stitch chain */
3510 			prev->m_next = m_free(m);
3511 			m = prev;
3512 			continue;
3513 		}
3514 		else if (m->m_next != NULL &&
3515 			     M_TRAILINGSPACE(m) >= shortfall &&
3516 			     m->m_next->m_len >= (8 + shortfall)) {
3517 		    /* m is writable and have enough data in next, pull up. */
3518 
3519 		  	memcpy(m->m_data + m->m_len, m->m_next->m_data,
3520 			    shortfall);
3521 			m->m_len += shortfall;
3522 			m->m_next->m_len -= shortfall;
3523 			m->m_next->m_data += shortfall;
3524 		}
3525 		else if (m->m_next == NULL || 1) {
3526 		  	/* Got a runt at the very end of the packet.
3527 			 * borrow data from the tail of the preceding mbuf and
3528 			 * update its length in-place. (The original data is still
3529 			 * valid, so we can do this even if prev is not writable.)
3530 			 */
3531 
3532 			/* if we'd make prev a runt, just move all of its data. */
3533 			KASSERT(prev != NULL /*, ("runt but null PREV")*/);
3534 			KASSERT(prev->m_len >= 8 /*, ("runt prev")*/);
3535 
3536 			if ((prev->m_len - shortfall) < 8)
3537 				shortfall = prev->m_len;
3538 
3539 #ifdef notyet	/* just do the safe slow thing for now */
3540 			if (!M_READONLY(m)) {
3541 				if (M_LEADINGSPACE(m) < shorfall) {
3542 					void *m_dat;
3543 					m_dat = (m->m_flags & M_PKTHDR) ?
3544 					  m->m_pktdat : m->dat;
3545 					memmove(m_dat, mtod(m, void*), m->m_len);
3546 					m->m_data = m_dat;
3547 				    }
3548 			} else
3549 #endif	/* just do the safe slow thing */
3550 			{
3551 				struct mbuf * n = NULL;
3552 				int newprevlen = prev->m_len - shortfall;
3553 
3554 				MGET(n, M_NOWAIT, MT_DATA);
3555 				if (n == NULL)
3556 				   return ENOBUFS;
3557 				KASSERT(m->m_len + shortfall < MLEN
3558 					/*,
3559 					  ("runt %d +prev %d too big\n", m->m_len, shortfall)*/);
3560 
3561 				/* first copy the data we're stealing from prev */
3562 				memcpy(n->m_data, prev->m_data + newprevlen,
3563 				    shortfall);
3564 
3565 				/* update prev->m_len accordingly */
3566 				prev->m_len -= shortfall;
3567 
3568 				/* copy data from runt m */
3569 				memcpy(n->m_data + shortfall, m->m_data,
3570 				    m->m_len);
3571 
3572 				/* n holds what we stole from prev, plus m */
3573 				n->m_len = shortfall + m->m_len;
3574 
3575 				/* stitch n into chain and free m */
3576 				n->m_next = m->m_next;
3577 				prev->m_next = n;
3578 				/* KASSERT(m->m_next == NULL); */
3579 				m->m_next = NULL;
3580 				m_free(m);
3581 				m = n;	/* for continuing loop */
3582 			}
3583 		}
3584 		prevlen = m->m_len;
3585 	}
3586 	return 0;
3587 }
3588 
3589 /*
3590  * Encapsulate an mbuf chain in the tx ring  by coupling the mbuf data
3591  * pointers to descriptors.
3592  */
3593 static int
3594 bge_encap(struct bge_softc *sc, struct mbuf *m_head, u_int32_t *txidx)
3595 {
3596 	struct bge_tx_bd	*f = NULL;
3597 	u_int32_t		frag, cur;
3598 	u_int16_t		csum_flags = 0;
3599 	u_int16_t		txbd_tso_flags = 0;
3600 	struct txdmamap_pool_entry *dma;
3601 	bus_dmamap_t dmamap;
3602 	int			i = 0;
3603 	struct m_tag		*mtag;
3604 	int			use_tso, maxsegsize, error;
3605 
3606 	cur = frag = *txidx;
3607 
3608 	if (m_head->m_pkthdr.csum_flags) {
3609 		if (m_head->m_pkthdr.csum_flags & M_CSUM_IPv4)
3610 			csum_flags |= BGE_TXBDFLAG_IP_CSUM;
3611 		if (m_head->m_pkthdr.csum_flags & (M_CSUM_TCPv4|M_CSUM_UDPv4))
3612 			csum_flags |= BGE_TXBDFLAG_TCP_UDP_CSUM;
3613 	}
3614 
3615 	/*
3616 	 * If we were asked to do an outboard checksum, and the NIC
3617 	 * has the bug where it sometimes adds in the Ethernet padding,
3618 	 * explicitly pad with zeros so the cksum will be correct either way.
3619 	 * (For now, do this for all chip versions, until newer
3620 	 * are confirmed to not require the workaround.)
3621 	 */
3622 	if ((csum_flags & BGE_TXBDFLAG_TCP_UDP_CSUM) == 0 ||
3623 #ifdef notyet
3624 	    (sc->bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||
3625 #endif
3626 	    m_head->m_pkthdr.len >= ETHER_MIN_NOPAD)
3627 		goto check_dma_bug;
3628 
3629 	if (bge_cksum_pad(m_head) != 0) {
3630 	    return ENOBUFS;
3631 	}
3632 
3633 check_dma_bug:
3634 	if (!(sc->bge_quirks & BGE_QUIRK_5700_SMALLDMA))
3635 		goto doit;
3636 	/*
3637 	 * bcm5700 Revision B silicon cannot handle DMA descriptors with
3638 	 * less than eight bytes.  If we encounter a teeny mbuf
3639 	 * at the end of a chain, we can pad.  Otherwise, copy.
3640 	 */
3641 	if (bge_compact_dma_runt(m_head) != 0)
3642 		return ENOBUFS;
3643 
3644 doit:
3645 	dma = SLIST_FIRST(&sc->txdma_list);
3646 	if (dma == NULL)
3647 		return ENOBUFS;
3648 	dmamap = dma->dmamap;
3649 
3650 	/*
3651 	 * Set up any necessary TSO state before we start packing...
3652 	 */
3653 	use_tso = (m_head->m_pkthdr.csum_flags & M_CSUM_TSOv4) != 0;
3654 	if (!use_tso) {
3655 		maxsegsize = 0;
3656 	} else {	/* TSO setup */
3657 		unsigned  mss;
3658 		struct ether_header *eh;
3659 		unsigned ip_tcp_hlen, iptcp_opt_words, tcp_seg_flags, offset;
3660 		struct mbuf * m0 = m_head;
3661 		struct ip *ip;
3662 		struct tcphdr *th;
3663 		int iphl, hlen;
3664 
3665 		/*
3666 		 * XXX It would be nice if the mbuf pkthdr had offset
3667 		 * fields for the protocol headers.
3668 		 */
3669 
3670 		eh = mtod(m0, struct ether_header *);
3671 		switch (htons(eh->ether_type)) {
3672 		case ETHERTYPE_IP:
3673 			offset = ETHER_HDR_LEN;
3674 			break;
3675 
3676 		case ETHERTYPE_VLAN:
3677 			offset = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3678 			break;
3679 
3680 		default:
3681 			/*
3682 			 * Don't support this protocol or encapsulation.
3683 			 */
3684 			return (ENOBUFS);
3685 		}
3686 
3687 		/*
3688 		 * TCP/IP headers are in the first mbuf; we can do
3689 		 * this the easy way.
3690 		 */
3691 		iphl = M_CSUM_DATA_IPv4_IPHL(m0->m_pkthdr.csum_data);
3692 		hlen = iphl + offset;
3693 		if (__predict_false(m0->m_len <
3694 				    (hlen + sizeof(struct tcphdr)))) {
3695 
3696 			aprint_debug_dev(sc->bge_dev,
3697 			    "TSO: hard case m0->m_len == %d < ip/tcp hlen %zd,"
3698 			    "not handled yet\n",
3699 			     m0->m_len, hlen+ sizeof(struct tcphdr));
3700 #ifdef NOTYET
3701 			/*
3702 			 * XXX jonathan@NetBSD.org: untested.
3703 			 * how to force  this branch to be taken?
3704 			 */
3705 			BGE_EVCNT_INCR(&sc->sc_ev_txtsopain);
3706 
3707 			m_copydata(m0, offset, sizeof(ip), &ip);
3708 			m_copydata(m0, hlen, sizeof(th), &th);
3709 
3710 			ip.ip_len = 0;
3711 
3712 			m_copyback(m0, hlen + offsetof(struct ip, ip_len),
3713 			    sizeof(ip.ip_len), &ip.ip_len);
3714 
3715 			th.th_sum = in_cksum_phdr(ip.ip_src.s_addr,
3716 			    ip.ip_dst.s_addr, htons(IPPROTO_TCP));
3717 
3718 			m_copyback(m0, hlen + offsetof(struct tcphdr, th_sum),
3719 			    sizeof(th.th_sum), &th.th_sum);
3720 
3721 			hlen += th.th_off << 2;
3722 			iptcp_opt_words	= hlen;
3723 #else
3724 			/*
3725 			 * if_wm "hard" case not yet supported, can we not
3726 			 * mandate it out of existence?
3727 			 */
3728 			(void) ip; (void)th; (void) ip_tcp_hlen;
3729 
3730 			return ENOBUFS;
3731 #endif
3732 		} else {
3733 			ip = (struct ip *) (mtod(m0, char *) + offset);
3734 			th = (struct tcphdr *) (mtod(m0, char *) + hlen);
3735 			ip_tcp_hlen = iphl +  (th->th_off << 2);
3736 
3737 			/* Total IP/TCP options, in 32-bit words */
3738 			iptcp_opt_words = (ip_tcp_hlen
3739 					   - sizeof(struct tcphdr)
3740 					   - sizeof(struct ip)) >> 2;
3741 		}
3742 		if (BGE_IS_5750_OR_BEYOND(sc)) {
3743 			th->th_sum = 0;
3744 			csum_flags &= ~(BGE_TXBDFLAG_TCP_UDP_CSUM);
3745 		} else {
3746 			/*
3747 			 * XXX jonathan@NetBSD.org: 5705 untested.
3748 			 * Requires TSO firmware patch for 5701/5703/5704.
3749 			 */
3750 			th->th_sum = in_cksum_phdr(ip->ip_src.s_addr,
3751 			    ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3752 		}
3753 
3754 		mss = m_head->m_pkthdr.segsz;
3755 		txbd_tso_flags |=
3756 		    BGE_TXBDFLAG_CPU_PRE_DMA |
3757 		    BGE_TXBDFLAG_CPU_POST_DMA;
3758 
3759 		/*
3760 		 * Our NIC TSO-assist assumes TSO has standard, optionless
3761 		 * IPv4 and TCP headers, which total 40 bytes. By default,
3762 		 * the NIC copies 40 bytes of IP/TCP header from the
3763 		 * supplied header into the IP/TCP header portion of
3764 		 * each post-TSO-segment. If the supplied packet has IP or
3765 		 * TCP options, we need to tell the NIC to copy those extra
3766 		 * bytes into each  post-TSO header, in addition to the normal
3767 		 * 40-byte IP/TCP header (and to leave space accordingly).
3768 		 * Unfortunately, the driver encoding of option length
3769 		 * varies across different ASIC families.
3770 		 */
3771 		tcp_seg_flags = 0;
3772 		if (iptcp_opt_words) {
3773 			if ( BGE_IS_5705_OR_BEYOND(sc)) {
3774 				tcp_seg_flags =
3775 					iptcp_opt_words << 11;
3776 			} else {
3777 				txbd_tso_flags |=
3778 					iptcp_opt_words << 12;
3779 			}
3780 		}
3781 		maxsegsize = mss | tcp_seg_flags;
3782 		ip->ip_len = htons(mss + ip_tcp_hlen);
3783 
3784 	}	/* TSO setup */
3785 
3786 	/*
3787 	 * Start packing the mbufs in this chain into
3788 	 * the fragment pointers. Stop when we run out
3789 	 * of fragments or hit the end of the mbuf chain.
3790 	 */
3791 	error = bus_dmamap_load_mbuf(sc->bge_dmatag, dmamap, m_head,
3792 	    BUS_DMA_NOWAIT);
3793 	if (error) {
3794 		return(ENOBUFS);
3795 	}
3796 	/*
3797 	 * Sanity check: avoid coming within 16 descriptors
3798 	 * of the end of the ring.
3799 	 */
3800 	if (dmamap->dm_nsegs > (BGE_TX_RING_CNT - sc->bge_txcnt - 16)) {
3801 		BGE_TSO_PRINTF(("%s: "
3802 		    " dmamap_load_mbuf too close to ring wrap\n",
3803 		    device_xname(sc->bge_dev)));
3804 		goto fail_unload;
3805 	}
3806 
3807 	mtag = sc->ethercom.ec_nvlans ?
3808 	    m_tag_find(m_head, PACKET_TAG_VLAN, NULL) : NULL;
3809 
3810 
3811 	/* Iterate over dmap-map fragments. */
3812 	for (i = 0; i < dmamap->dm_nsegs; i++) {
3813 		f = &sc->bge_rdata->bge_tx_ring[frag];
3814 		if (sc->bge_cdata.bge_tx_chain[frag] != NULL)
3815 			break;
3816 
3817 		bge_set_hostaddr(&f->bge_addr, dmamap->dm_segs[i].ds_addr);
3818 		f->bge_len = dmamap->dm_segs[i].ds_len;
3819 
3820 		/*
3821 		 * For 5751 and follow-ons, for TSO we must turn
3822 		 * off checksum-assist flag in the tx-descr, and
3823 		 * supply the ASIC-revision-specific encoding
3824 		 * of TSO flags and segsize.
3825 		 */
3826 		if (use_tso) {
3827 			if (BGE_IS_5750_OR_BEYOND(sc) || i == 0) {
3828 				f->bge_rsvd = maxsegsize;
3829 				f->bge_flags = csum_flags | txbd_tso_flags;
3830 			} else {
3831 				f->bge_rsvd = 0;
3832 				f->bge_flags =
3833 				  (csum_flags | txbd_tso_flags) & 0x0fff;
3834 			}
3835 		} else {
3836 			f->bge_rsvd = 0;
3837 			f->bge_flags = csum_flags;
3838 		}
3839 
3840 		if (mtag != NULL) {
3841 			f->bge_flags |= BGE_TXBDFLAG_VLAN_TAG;
3842 			f->bge_vlan_tag = VLAN_TAG_VALUE(mtag);
3843 		} else {
3844 			f->bge_vlan_tag = 0;
3845 		}
3846 		cur = frag;
3847 		BGE_INC(frag, BGE_TX_RING_CNT);
3848 	}
3849 
3850 	if (i < dmamap->dm_nsegs) {
3851 		BGE_TSO_PRINTF(("%s: reached %d < dm_nsegs %d\n",
3852 		    device_xname(sc->bge_dev), i, dmamap->dm_nsegs));
3853 		goto fail_unload;
3854 	}
3855 
3856 	bus_dmamap_sync(sc->bge_dmatag, dmamap, 0, dmamap->dm_mapsize,
3857 	    BUS_DMASYNC_PREWRITE);
3858 
3859 	if (frag == sc->bge_tx_saved_considx) {
3860 		BGE_TSO_PRINTF(("%s: frag %d = wrapped id %d?\n",
3861 		    device_xname(sc->bge_dev), frag, sc->bge_tx_saved_considx));
3862 
3863 		goto fail_unload;
3864 	}
3865 
3866 	sc->bge_rdata->bge_tx_ring[cur].bge_flags |= BGE_TXBDFLAG_END;
3867 	sc->bge_cdata.bge_tx_chain[cur] = m_head;
3868 	SLIST_REMOVE_HEAD(&sc->txdma_list, link);
3869 	sc->txdma[cur] = dma;
3870 	sc->bge_txcnt += dmamap->dm_nsegs;
3871 
3872 	*txidx = frag;
3873 
3874 	return(0);
3875 
3876  fail_unload:
3877 	bus_dmamap_unload(sc->bge_dmatag, dmamap);
3878 
3879 	return ENOBUFS;
3880 }
3881 
3882 /*
3883  * Main transmit routine. To avoid having to do mbuf copies, we put pointers
3884  * to the mbuf data regions directly in the transmit descriptors.
3885  */
3886 static void
3887 bge_start(struct ifnet *ifp)
3888 {
3889 	struct bge_softc *sc;
3890 	struct mbuf *m_head = NULL;
3891 	u_int32_t prodidx;
3892 	int pkts = 0;
3893 
3894 	sc = ifp->if_softc;
3895 
3896 	if ((ifp->if_flags & (IFF_RUNNING|IFF_OACTIVE)) != IFF_RUNNING)
3897 		return;
3898 
3899 	prodidx = sc->bge_tx_prodidx;
3900 
3901 	while(sc->bge_cdata.bge_tx_chain[prodidx] == NULL) {
3902 		IFQ_POLL(&ifp->if_snd, m_head);
3903 		if (m_head == NULL)
3904 			break;
3905 
3906 #if 0
3907 		/*
3908 		 * XXX
3909 		 * safety overkill.  If this is a fragmented packet chain
3910 		 * with delayed TCP/UDP checksums, then only encapsulate
3911 		 * it if we have enough descriptors to handle the entire
3912 		 * chain at once.
3913 		 * (paranoia -- may not actually be needed)
3914 		 */
3915 		if (m_head->m_flags & M_FIRSTFRAG &&
3916 		    m_head->m_pkthdr.csum_flags & (CSUM_DELAY_DATA)) {
3917 			if ((BGE_TX_RING_CNT - sc->bge_txcnt) <
3918 			    M_CSUM_DATA_IPv4_OFFSET(m_head->m_pkthdr.csum_data) + 16) {
3919 				ifp->if_flags |= IFF_OACTIVE;
3920 				break;
3921 			}
3922 		}
3923 #endif
3924 
3925 		/*
3926 		 * Pack the data into the transmit ring. If we
3927 		 * don't have room, set the OACTIVE flag and wait
3928 		 * for the NIC to drain the ring.
3929 		 */
3930 		if (bge_encap(sc, m_head, &prodidx)) {
3931 			ifp->if_flags |= IFF_OACTIVE;
3932 			break;
3933 		}
3934 
3935 		/* now we are committed to transmit the packet */
3936 		IFQ_DEQUEUE(&ifp->if_snd, m_head);
3937 		pkts++;
3938 
3939 #if NBPFILTER > 0
3940 		/*
3941 		 * If there's a BPF listener, bounce a copy of this frame
3942 		 * to him.
3943 		 */
3944 		if (ifp->if_bpf)
3945 			bpf_mtap(ifp->if_bpf, m_head);
3946 #endif
3947 	}
3948 	if (pkts == 0)
3949 		return;
3950 
3951 	/* Transmit */
3952 	CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
3953 	if (sc->bge_quirks & BGE_QUIRK_PRODUCER_BUG)	/* 5700 b2 errata */
3954 		CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
3955 
3956 	sc->bge_tx_prodidx = prodidx;
3957 
3958 	/*
3959 	 * Set a timeout in case the chip goes out to lunch.
3960 	 */
3961 	ifp->if_timer = 5;
3962 }
3963 
3964 static int
3965 bge_init(struct ifnet *ifp)
3966 {
3967 	struct bge_softc *sc = ifp->if_softc;
3968 	const u_int16_t *m;
3969 	int s, error = 0;
3970 
3971 	s = splnet();
3972 
3973 	ifp = &sc->ethercom.ec_if;
3974 
3975 	/* Cancel pending I/O and flush buffers. */
3976 	bge_stop(ifp, 0);
3977 	bge_reset(sc);
3978 	bge_chipinit(sc);
3979 
3980 	/*
3981 	 * Init the various state machines, ring
3982 	 * control blocks and firmware.
3983 	 */
3984 	error = bge_blockinit(sc);
3985 	if (error != 0) {
3986 		aprint_error_dev(sc->bge_dev, "initialization error %d\n",
3987 		    error);
3988 		splx(s);
3989 		return error;
3990 	}
3991 
3992 	ifp = &sc->ethercom.ec_if;
3993 
3994 	/* Specify MTU. */
3995 	CSR_WRITE_4(sc, BGE_RX_MTU, ifp->if_mtu +
3996 	    ETHER_HDR_LEN + ETHER_CRC_LEN + ETHER_VLAN_ENCAP_LEN);
3997 
3998 	/* Load our MAC address. */
3999 	m = (const u_int16_t *)&(CLLADDR(ifp->if_sadl)[0]);
4000 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_LO, htons(m[0]));
4001 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_HI, (htons(m[1]) << 16) | htons(m[2]));
4002 
4003 	/* Enable or disable promiscuous mode as needed. */
4004 	if (ifp->if_flags & IFF_PROMISC) {
4005 		BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
4006 	} else {
4007 		BGE_CLRBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
4008 	}
4009 
4010 	/* Program multicast filter. */
4011 	bge_setmulti(sc);
4012 
4013 	/* Init RX ring. */
4014 	bge_init_rx_ring_std(sc);
4015 
4016 	/* Init jumbo RX ring. */
4017 	if (ifp->if_mtu > (ETHERMTU + ETHER_HDR_LEN + ETHER_CRC_LEN))
4018 		bge_init_rx_ring_jumbo(sc);
4019 
4020 	/* Init our RX return ring index */
4021 	sc->bge_rx_saved_considx = 0;
4022 
4023 	/* Init TX ring. */
4024 	bge_init_tx_ring(sc);
4025 
4026 	/* Turn on transmitter */
4027 	BGE_SETBIT(sc, BGE_TX_MODE, BGE_TXMODE_ENABLE);
4028 
4029 	/* Turn on receiver */
4030 	BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
4031 
4032 	CSR_WRITE_4(sc, BGE_MAX_RX_FRAME_LOWAT, 2);
4033 
4034 	/* Tell firmware we're alive. */
4035 	BGE_SETBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
4036 
4037 	/* Enable host interrupts. */
4038 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_CLEAR_INTA);
4039 	BGE_CLRBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
4040 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 0);
4041 
4042 	if ((error = bge_ifmedia_upd(ifp)) != 0)
4043 		goto out;
4044 
4045 	ifp->if_flags |= IFF_RUNNING;
4046 	ifp->if_flags &= ~IFF_OACTIVE;
4047 
4048 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
4049 
4050 out:
4051 	splx(s);
4052 
4053 	return error;
4054 }
4055 
4056 /*
4057  * Set media options.
4058  */
4059 static int
4060 bge_ifmedia_upd(struct ifnet *ifp)
4061 {
4062 	struct bge_softc *sc = ifp->if_softc;
4063 	struct mii_data *mii = &sc->bge_mii;
4064 	struct ifmedia *ifm = &sc->bge_ifmedia;
4065 	int rc;
4066 
4067 	/* If this is a 1000baseX NIC, enable the TBI port. */
4068 	if (sc->bge_tbi) {
4069 		if (IFM_TYPE(ifm->ifm_media) != IFM_ETHER)
4070 			return(EINVAL);
4071 		switch(IFM_SUBTYPE(ifm->ifm_media)) {
4072 		case IFM_AUTO:
4073 			break;
4074 		case IFM_1000_SX:
4075 			if ((ifm->ifm_media & IFM_GMASK) == IFM_FDX) {
4076 				BGE_CLRBIT(sc, BGE_MAC_MODE,
4077 				    BGE_MACMODE_HALF_DUPLEX);
4078 			} else {
4079 				BGE_SETBIT(sc, BGE_MAC_MODE,
4080 				    BGE_MACMODE_HALF_DUPLEX);
4081 			}
4082 			break;
4083 		default:
4084 			return(EINVAL);
4085 		}
4086 		/* XXX 802.3x flow control for 1000BASE-SX */
4087 		return(0);
4088 	}
4089 
4090 	sc->bge_link = 0;
4091 	if ((rc = mii_mediachg(mii)) == ENXIO)
4092 		return 0;
4093 	return rc;
4094 }
4095 
4096 /*
4097  * Report current media status.
4098  */
4099 static void
4100 bge_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
4101 {
4102 	struct bge_softc *sc = ifp->if_softc;
4103 	struct mii_data *mii = &sc->bge_mii;
4104 
4105 	if (sc->bge_tbi) {
4106 		ifmr->ifm_status = IFM_AVALID;
4107 		ifmr->ifm_active = IFM_ETHER;
4108 		if (CSR_READ_4(sc, BGE_MAC_STS) &
4109 		    BGE_MACSTAT_TBI_PCS_SYNCHED)
4110 			ifmr->ifm_status |= IFM_ACTIVE;
4111 		ifmr->ifm_active |= IFM_1000_SX;
4112 		if (CSR_READ_4(sc, BGE_MAC_MODE) & BGE_MACMODE_HALF_DUPLEX)
4113 			ifmr->ifm_active |= IFM_HDX;
4114 		else
4115 			ifmr->ifm_active |= IFM_FDX;
4116 		return;
4117 	}
4118 
4119 	mii_pollstat(mii);
4120 	ifmr->ifm_status = mii->mii_media_status;
4121 	ifmr->ifm_active = (mii->mii_media_active & ~IFM_ETH_FMASK) |
4122 	    sc->bge_flowflags;
4123 }
4124 
4125 static int
4126 bge_ioctl(struct ifnet *ifp, u_long command, void *data)
4127 {
4128 	struct bge_softc *sc = ifp->if_softc;
4129 	struct ifreq *ifr = (struct ifreq *) data;
4130 	int s, error = 0;
4131 	struct mii_data *mii;
4132 
4133 	s = splnet();
4134 
4135 	switch(command) {
4136 	case SIOCSIFFLAGS:
4137 		if (ifp->if_flags & IFF_UP) {
4138 			/*
4139 			 * If only the state of the PROMISC flag changed,
4140 			 * then just use the 'set promisc mode' command
4141 			 * instead of reinitializing the entire NIC. Doing
4142 			 * a full re-init means reloading the firmware and
4143 			 * waiting for it to start up, which may take a
4144 			 * second or two.
4145 			 */
4146 			if (ifp->if_flags & IFF_RUNNING &&
4147 			    ifp->if_flags & IFF_PROMISC &&
4148 			    !(sc->bge_if_flags & IFF_PROMISC)) {
4149 				BGE_SETBIT(sc, BGE_RX_MODE,
4150 				    BGE_RXMODE_RX_PROMISC);
4151 			} else if (ifp->if_flags & IFF_RUNNING &&
4152 			    !(ifp->if_flags & IFF_PROMISC) &&
4153 			    sc->bge_if_flags & IFF_PROMISC) {
4154 				BGE_CLRBIT(sc, BGE_RX_MODE,
4155 				    BGE_RXMODE_RX_PROMISC);
4156 			} else if (!(sc->bge_if_flags & IFF_UP))
4157 				bge_init(ifp);
4158 		} else {
4159 			if (ifp->if_flags & IFF_RUNNING)
4160 				bge_stop(ifp, 1);
4161 		}
4162 		sc->bge_if_flags = ifp->if_flags;
4163 		error = 0;
4164 		break;
4165 	case SIOCSIFMEDIA:
4166 		/* XXX Flow control is not supported for 1000BASE-SX */
4167 		if (sc->bge_tbi) {
4168 			ifr->ifr_media &= ~IFM_ETH_FMASK;
4169 			sc->bge_flowflags = 0;
4170 		}
4171 
4172 		/* Flow control requires full-duplex mode. */
4173 		if (IFM_SUBTYPE(ifr->ifr_media) == IFM_AUTO ||
4174 		    (ifr->ifr_media & IFM_FDX) == 0) {
4175 		    	ifr->ifr_media &= ~IFM_ETH_FMASK;
4176 		}
4177 		if (IFM_SUBTYPE(ifr->ifr_media) != IFM_AUTO) {
4178 			if ((ifr->ifr_media & IFM_ETH_FMASK) == IFM_FLOW) {
4179 				/* We an do both TXPAUSE and RXPAUSE. */
4180 				ifr->ifr_media |=
4181 				    IFM_ETH_TXPAUSE | IFM_ETH_RXPAUSE;
4182 			}
4183 			sc->bge_flowflags = ifr->ifr_media & IFM_ETH_FMASK;
4184 		}
4185 		/* FALLTHROUGH */
4186 	case SIOCGIFMEDIA:
4187 		if (sc->bge_tbi) {
4188 			error = ifmedia_ioctl(ifp, ifr, &sc->bge_ifmedia,
4189 			    command);
4190 		} else {
4191 			mii = &sc->bge_mii;
4192 			error = ifmedia_ioctl(ifp, ifr, &mii->mii_media,
4193 			    command);
4194 		}
4195 		break;
4196 	default:
4197 		if ((error = ether_ioctl(ifp, command, data)) != ENETRESET)
4198 			break;
4199 
4200 		error = 0;
4201 
4202 		if (command != SIOCADDMULTI && command != SIOCDELMULTI)
4203 			;
4204 		else if (ifp->if_flags & IFF_RUNNING)
4205 			bge_setmulti(sc);
4206 		break;
4207 	}
4208 
4209 	splx(s);
4210 
4211 	return(error);
4212 }
4213 
4214 static void
4215 bge_watchdog(struct ifnet *ifp)
4216 {
4217 	struct bge_softc *sc;
4218 
4219 	sc = ifp->if_softc;
4220 
4221 	aprint_error_dev(sc->bge_dev, "watchdog timeout -- resetting\n");
4222 
4223 	ifp->if_flags &= ~IFF_RUNNING;
4224 	bge_init(ifp);
4225 
4226 	ifp->if_oerrors++;
4227 }
4228 
4229 static void
4230 bge_stop_block(struct bge_softc *sc, bus_addr_t reg, uint32_t bit)
4231 {
4232 	int i;
4233 
4234 	BGE_CLRBIT(sc, reg, bit);
4235 
4236 	for (i = 0; i < BGE_TIMEOUT; i++) {
4237 		if ((CSR_READ_4(sc, reg) & bit) == 0)
4238 			return;
4239 		delay(100);
4240 		if (sc->bge_pcie)
4241 		  DELAY(1000);
4242 	}
4243 
4244 	aprint_error_dev(sc->bge_dev,
4245 	    "block failed to stop: reg 0x%lx, bit 0x%08x\n", (u_long)reg, bit);
4246 }
4247 
4248 /*
4249  * Stop the adapter and free any mbufs allocated to the
4250  * RX and TX lists.
4251  */
4252 static void
4253 bge_stop(struct ifnet *ifp, int disable)
4254 {
4255 	struct bge_softc *sc = ifp->if_softc;
4256 
4257 	callout_stop(&sc->bge_timeout);
4258 
4259 	/*
4260 	 * Disable all of the receiver blocks
4261 	 */
4262 	bge_stop_block(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
4263 	bge_stop_block(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
4264 	bge_stop_block(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
4265 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4266 		bge_stop_block(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
4267 	}
4268 	bge_stop_block(sc, BGE_RDBDI_MODE, BGE_RBDIMODE_ENABLE);
4269 	bge_stop_block(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
4270 	bge_stop_block(sc, BGE_RBDC_MODE, BGE_RBDCMODE_ENABLE);
4271 
4272 	/*
4273 	 * Disable all of the transmit blocks
4274 	 */
4275 	bge_stop_block(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
4276 	bge_stop_block(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
4277 	bge_stop_block(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
4278 	bge_stop_block(sc, BGE_RDMA_MODE, BGE_RDMAMODE_ENABLE);
4279 	bge_stop_block(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
4280 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4281 		bge_stop_block(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
4282 	}
4283 	bge_stop_block(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
4284 
4285 	/*
4286 	 * Shut down all of the memory managers and related
4287 	 * state machines.
4288 	 */
4289 	bge_stop_block(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
4290 	bge_stop_block(sc, BGE_WDMA_MODE, BGE_WDMAMODE_ENABLE);
4291 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4292 		bge_stop_block(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
4293 	}
4294 
4295 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0xFFFFFFFF);
4296 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0);
4297 
4298 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4299 		bge_stop_block(sc, BGE_BMAN_MODE, BGE_BMANMODE_ENABLE);
4300 		bge_stop_block(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE);
4301 	}
4302 
4303 	/* Disable host interrupts. */
4304 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
4305 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 1);
4306 
4307 	/*
4308 	 * Tell firmware we're shutting down.
4309 	 */
4310 	BGE_CLRBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
4311 
4312 	/* Free the RX lists. */
4313 	bge_free_rx_ring_std(sc);
4314 
4315 	/* Free jumbo RX list. */
4316 	bge_free_rx_ring_jumbo(sc);
4317 
4318 	/* Free TX buffers. */
4319 	bge_free_tx_ring(sc);
4320 
4321 	/*
4322 	 * Isolate/power down the PHY.
4323 	 */
4324 	if (!sc->bge_tbi)
4325 		mii_down(&sc->bge_mii);
4326 
4327 	sc->bge_link = 0;
4328 
4329 	sc->bge_tx_saved_considx = BGE_TXCONS_UNSET;
4330 
4331 	ifp->if_flags &= ~(IFF_RUNNING | IFF_OACTIVE);
4332 }
4333 
4334 static int
4335 sysctl_bge_verify(SYSCTLFN_ARGS)
4336 {
4337 	int error, t;
4338 	struct sysctlnode node;
4339 
4340 	node = *rnode;
4341 	t = *(int*)rnode->sysctl_data;
4342 	node.sysctl_data = &t;
4343 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
4344 	if (error || newp == NULL)
4345 		return (error);
4346 
4347 #if 0
4348 	DPRINTF2(("%s: t = %d, nodenum = %d, rnodenum = %d\n", __func__, t,
4349 	    node.sysctl_num, rnode->sysctl_num));
4350 #endif
4351 
4352 	if (node.sysctl_num == bge_rxthresh_nodenum) {
4353 		if (t < 0 || t >= NBGE_RX_THRESH)
4354 			return (EINVAL);
4355 		bge_update_all_threshes(t);
4356 	} else
4357 		return (EINVAL);
4358 
4359 	*(int*)rnode->sysctl_data = t;
4360 
4361 	return (0);
4362 }
4363 
4364 /*
4365  * Set up sysctl(3) MIB, hw.bge.*.
4366  *
4367  * TBD condition SYSCTL_PERMANENT on being an LKM or not
4368  */
4369 SYSCTL_SETUP(sysctl_bge, "sysctl bge subtree setup")
4370 {
4371 	int rc, bge_root_num;
4372 	const struct sysctlnode *node;
4373 
4374 	if ((rc = sysctl_createv(clog, 0, NULL, NULL,
4375 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "hw", NULL,
4376 	    NULL, 0, NULL, 0, CTL_HW, CTL_EOL)) != 0) {
4377 		goto err;
4378 	}
4379 
4380 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
4381 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "bge",
4382 	    SYSCTL_DESCR("BGE interface controls"),
4383 	    NULL, 0, NULL, 0, CTL_HW, CTL_CREATE, CTL_EOL)) != 0) {
4384 		goto err;
4385 	}
4386 
4387 	bge_root_num = node->sysctl_num;
4388 
4389 	/* BGE Rx interrupt mitigation level */
4390 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
4391 	    CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
4392 	    CTLTYPE_INT, "rx_lvl",
4393 	    SYSCTL_DESCR("BGE receive interrupt mitigation level"),
4394 	    sysctl_bge_verify, 0,
4395 	    &bge_rx_thresh_lvl,
4396 	    0, CTL_HW, bge_root_num, CTL_CREATE,
4397 	    CTL_EOL)) != 0) {
4398 		goto err;
4399 	}
4400 
4401 	bge_rxthresh_nodenum = node->sysctl_num;
4402 
4403 	return;
4404 
4405 err:
4406 	aprint_error("%s: sysctl_createv failed (rc = %d)\n", __func__, rc);
4407 }
4408