xref: /netbsd-src/sys/dev/pci/if_bge.c (revision d48f14661dda8638fee055ba15d35bdfb29b9fa8)
1 /*	$NetBSD: if_bge.c,v 1.109 2006/06/01 02:20:54 jonathan 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 1.109 2006/06/01 02:20:54 jonathan Exp $");
83 
84 #include "bpfilter.h"
85 #include "vlan.h"
86 
87 #include <sys/param.h>
88 #include <sys/systm.h>
89 #include <sys/callout.h>
90 #include <sys/sockio.h>
91 #include <sys/mbuf.h>
92 #include <sys/malloc.h>
93 #include <sys/kernel.h>
94 #include <sys/device.h>
95 #include <sys/socket.h>
96 #include <sys/sysctl.h>
97 
98 #include <net/if.h>
99 #include <net/if_dl.h>
100 #include <net/if_media.h>
101 #include <net/if_ether.h>
102 
103 #ifdef INET
104 #include <netinet/in.h>
105 #include <netinet/in_systm.h>
106 #include <netinet/in_var.h>
107 #include <netinet/ip.h>
108 #endif
109 
110 /* Headers for TCP  Segmentation Offload (TSO) */
111 #include <netinet/in_systm.h>		/* n_time for <netinet/ip.h>... */
112 #include <netinet/in.h>			/* ip_{src,dst}, for <netinet/ip.h> */
113 #include <netinet/ip.h>			/* for struct ip */
114 #include <netinet/tcp.h>		/* for struct tcphdr */
115 
116 
117 #if NBPFILTER > 0
118 #include <net/bpf.h>
119 #endif
120 
121 #include <dev/pci/pcireg.h>
122 #include <dev/pci/pcivar.h>
123 #include <dev/pci/pcidevs.h>
124 
125 #include <dev/mii/mii.h>
126 #include <dev/mii/miivar.h>
127 #include <dev/mii/miidevs.h>
128 #include <dev/mii/brgphyreg.h>
129 
130 #include <dev/pci/if_bgereg.h>
131 
132 #include <uvm/uvm_extern.h>
133 
134 #define ETHER_MIN_NOPAD (ETHER_MIN_LEN - ETHER_CRC_LEN) /* i.e., 60 */
135 
136 
137 /*
138  * Tunable thresholds for rx-side bge interrupt mitigation.
139  */
140 
141 /*
142  * The pairs of values below were obtained from empirical measurement
143  * on bcm5700 rev B2; they ar designed to give roughly 1 receive
144  * interrupt for every N packets received, where N is, approximately,
145  * the second value (rx_max_bds) in each pair.  The values are chosen
146  * such that moving from one pair to the succeeding pair was observed
147  * to roughly halve interrupt rate under sustained input packet load.
148  * The values were empirically chosen to avoid overflowing internal
149  * limits on the  bcm5700: inreasing rx_ticks much beyond 600
150  * results in internal wrapping and higher interrupt rates.
151  * The limit of 46 frames was chosen to match NFS workloads.
152  *
153  * These values also work well on bcm5701, bcm5704C, and (less
154  * tested) bcm5703.  On other chipsets, (including the Altima chip
155  * family), the larger values may overflow internal chip limits,
156  * leading to increasing interrupt rates rather than lower interrupt
157  * rates.
158  *
159  * Applications using heavy interrupt mitigation (interrupting every
160  * 32 or 46 frames) in both directions may need to increase the TCP
161  * windowsize to above 131072 bytes (e.g., to 199608 bytes) to sustain
162  * full link bandwidth, due to ACKs and window updates lingering
163  * in the RX queue during the 30-to-40-frame interrupt-mitigation window.
164  */
165 static const struct bge_load_rx_thresh {
166 	int rx_ticks;
167 	int rx_max_bds; }
168 bge_rx_threshes[] = {
169 	{ 32,   2 },
170 	{ 50,   4 },
171 	{ 100,  8 },
172 	{ 192, 16 },
173 	{ 416, 32 },
174 	{ 598, 46 }
175 };
176 #define NBGE_RX_THRESH (sizeof(bge_rx_threshes) / sizeof(bge_rx_threshes[0]))
177 
178 /* XXX patchable; should be sysctl'able */
179 static int	bge_auto_thresh = 1;
180 static int	bge_rx_thresh_lvl;
181 
182 static int	bge_rxthresh_nodenum;
183 
184 static int	bge_probe(device_t, cfdata_t, void *);
185 static void	bge_attach(device_t, device_t, void *);
186 static void	bge_powerhook(int, void *);
187 static void	bge_release_resources(struct bge_softc *);
188 static void	bge_txeof(struct bge_softc *);
189 static void	bge_rxeof(struct bge_softc *);
190 
191 static void	bge_tick(void *);
192 static void	bge_stats_update(struct bge_softc *);
193 static int	bge_encap(struct bge_softc *, struct mbuf *, u_int32_t *);
194 
195 static int	bge_intr(void *);
196 static void	bge_start(struct ifnet *);
197 static int	bge_ioctl(struct ifnet *, u_long, caddr_t);
198 static int	bge_init(struct ifnet *);
199 static void	bge_stop(struct bge_softc *);
200 static void	bge_watchdog(struct ifnet *);
201 static void	bge_shutdown(void *);
202 static int	bge_ifmedia_upd(struct ifnet *);
203 static void	bge_ifmedia_sts(struct ifnet *, struct ifmediareq *);
204 
205 static void	bge_setmulti(struct bge_softc *);
206 
207 static void	bge_handle_events(struct bge_softc *);
208 static int	bge_alloc_jumbo_mem(struct bge_softc *);
209 #if 0 /* XXX */
210 static void	bge_free_jumbo_mem(struct bge_softc *);
211 #endif
212 static void	*bge_jalloc(struct bge_softc *);
213 static void	bge_jfree(struct mbuf *, caddr_t, size_t, void *);
214 static int	bge_newbuf_std(struct bge_softc *, int, struct mbuf *,
215 			       bus_dmamap_t);
216 static int	bge_newbuf_jumbo(struct bge_softc *, int, struct mbuf *);
217 static int	bge_init_rx_ring_std(struct bge_softc *);
218 static void	bge_free_rx_ring_std(struct bge_softc *);
219 static int	bge_init_rx_ring_jumbo(struct bge_softc *);
220 static void	bge_free_rx_ring_jumbo(struct bge_softc *);
221 static void	bge_free_tx_ring(struct bge_softc *);
222 static int	bge_init_tx_ring(struct bge_softc *);
223 
224 static int	bge_chipinit(struct bge_softc *);
225 static int	bge_blockinit(struct bge_softc *);
226 static int	bge_setpowerstate(struct bge_softc *, int);
227 
228 static void	bge_reset(struct bge_softc *);
229 
230 #define BGE_DEBUG
231 #ifdef BGE_DEBUG
232 #define DPRINTF(x)	if (bgedebug) printf x
233 #define DPRINTFN(n,x)	if (bgedebug >= (n)) printf x
234 #define BGE_TSO_PRINTF(x)  do { if (bge_tso_debug) printf x ;} while (0)
235 int	bgedebug = 0;
236 int	bge_tso_debug = 0;
237 #else
238 #define DPRINTF(x)
239 #define DPRINTFN(n,x)
240 #define BGE_TSO_PRINTF(x)
241 #endif
242 
243 #ifdef BGE_EVENT_COUNTERS
244 #define	BGE_EVCNT_INCR(ev)	(ev).ev_count++
245 #define	BGE_EVCNT_ADD(ev, val)	(ev).ev_count += (val)
246 #define	BGE_EVCNT_UPD(ev, val)	(ev).ev_count = (val)
247 #else
248 #define	BGE_EVCNT_INCR(ev)	/* nothing */
249 #define	BGE_EVCNT_ADD(ev, val)	/* nothing */
250 #define	BGE_EVCNT_UPD(ev, val)	/* nothing */
251 #endif
252 
253 /* Various chip quirks. */
254 #define	BGE_QUIRK_LINK_STATE_BROKEN	0x00000001
255 #define	BGE_QUIRK_CSUM_BROKEN		0x00000002
256 #define	BGE_QUIRK_ONLY_PHY_1		0x00000004
257 #define	BGE_QUIRK_5700_SMALLDMA		0x00000008
258 #define	BGE_QUIRK_5700_PCIX_REG_BUG	0x00000010
259 #define	BGE_QUIRK_PRODUCER_BUG		0x00000020
260 #define	BGE_QUIRK_PCIX_DMA_ALIGN_BUG	0x00000040
261 #define	BGE_QUIRK_5705_CORE		0x00000080
262 #define	BGE_QUIRK_FEWER_MBUFS		0x00000100
263 
264 /*
265  * XXX: how to handle variants based on 5750 and derivatives:
266  * 5750 5751, 5721, possibly 5714, 5752, and 5708?, which
267  * in general behave like a 5705, except with additional quirks.
268  * This driver's current handling of the 5721 is wrong;
269  * how we map ASIC revision to "quirks" needs more thought.
270  * (defined here until the thought is done).
271  */
272 #define BGE_IS_5714_FAMILY(sc) \
273 	(BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5714 ||	\
274 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5780 ||	\
275 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5715 )
276 
277 #define BGE_IS_5750_OR_BEYOND(sc)  \
278 	(BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5750 || \
279 	 BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5752 || \
280 	 BGE_IS_5714_FAMILY(sc) )
281 
282 #define BGE_IS_5705_OR_BEYOND(sc)  \
283 	( ((sc)->bge_quirks & BGE_QUIRK_5705_CORE) || \
284 	  BGE_IS_5750_OR_BEYOND(sc) )
285 
286 
287 /* following bugs are common to bcm5700 rev B, all flavours */
288 #define BGE_QUIRK_5700_COMMON \
289 	(BGE_QUIRK_5700_SMALLDMA|BGE_QUIRK_PRODUCER_BUG)
290 
291 CFATTACH_DECL(bge, sizeof(struct bge_softc),
292     bge_probe, bge_attach, NULL, NULL);
293 
294 static u_int32_t
295 bge_readmem_ind(struct bge_softc *sc, int off)
296 {
297 	struct pci_attach_args	*pa = &(sc->bge_pa);
298 	pcireg_t val;
299 
300 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_BASEADDR, off);
301 	val = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_DATA);
302 	return val;
303 }
304 
305 static void
306 bge_writemem_ind(struct bge_softc *sc, int off, int val)
307 {
308 	struct pci_attach_args	*pa = &(sc->bge_pa);
309 
310 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_BASEADDR, off);
311 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MEMWIN_DATA, val);
312 }
313 
314 #ifdef notdef
315 static u_int32_t
316 bge_readreg_ind(struct bge_softc *sc, int off)
317 {
318 	struct pci_attach_args	*pa = &(sc->bge_pa);
319 
320 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_REG_BASEADDR, off);
321 	return(pci_conf_read(pa->pa_pc, pa->pa_tag, 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 	struct pci_attach_args	*pa = &(sc->bge_pa);
329 
330 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_REG_BASEADDR, off);
331 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_REG_DATA, val);
332 }
333 
334 #ifdef notdef
335 static u_int8_t
336 bge_vpd_readbyte(struct bge_softc *sc, int addr)
337 {
338 	int i;
339 	u_int32_t val;
340 	struct pci_attach_args	*pa = &(sc->bge_pa);
341 
342 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_VPD_ADDR, addr);
343 	for (i = 0; i < BGE_TIMEOUT * 10; i++) {
344 		DELAY(10);
345 		if (pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_VPD_ADDR) &
346 		    BGE_VPD_FLAG)
347 			break;
348 	}
349 
350 	if (i == BGE_TIMEOUT) {
351 		printf("%s: VPD read timed out\n", sc->bge_dev.dv_xname);
352 		return(0);
353 	}
354 
355 	val = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_VPD_DATA);
356 
357 	return((val >> ((addr % 4) * 8)) & 0xFF);
358 }
359 
360 static void
361 bge_vpd_read_res(struct bge_softc *sc, struct vpd_res *res, int addr)
362 {
363 	int i;
364 	u_int8_t *ptr;
365 
366 	ptr = (u_int8_t *)res;
367 	for (i = 0; i < sizeof(struct vpd_res); i++)
368 		ptr[i] = bge_vpd_readbyte(sc, i + addr);
369 }
370 
371 static void
372 bge_vpd_read(struct bge_softc *sc)
373 {
374 	int pos = 0, i;
375 	struct vpd_res res;
376 
377 	if (sc->bge_vpd_prodname != NULL)
378 		free(sc->bge_vpd_prodname, M_DEVBUF);
379 	if (sc->bge_vpd_readonly != NULL)
380 		free(sc->bge_vpd_readonly, M_DEVBUF);
381 	sc->bge_vpd_prodname = NULL;
382 	sc->bge_vpd_readonly = NULL;
383 
384 	bge_vpd_read_res(sc, &res, pos);
385 
386 	if (res.vr_id != VPD_RES_ID) {
387 		printf("%s: bad VPD resource id: expected %x got %x\n",
388 			sc->bge_dev.dv_xname, VPD_RES_ID, res.vr_id);
389 		return;
390 	}
391 
392 	pos += sizeof(res);
393 	sc->bge_vpd_prodname = malloc(res.vr_len + 1, M_DEVBUF, M_NOWAIT);
394 	if (sc->bge_vpd_prodname == NULL)
395 		panic("bge_vpd_read");
396 	for (i = 0; i < res.vr_len; i++)
397 		sc->bge_vpd_prodname[i] = bge_vpd_readbyte(sc, i + pos);
398 	sc->bge_vpd_prodname[i] = '\0';
399 	pos += i;
400 
401 	bge_vpd_read_res(sc, &res, pos);
402 
403 	if (res.vr_id != VPD_RES_READ) {
404 		printf("%s: bad VPD resource id: expected %x got %x\n",
405 		    sc->bge_dev.dv_xname, VPD_RES_READ, res.vr_id);
406 		return;
407 	}
408 
409 	pos += sizeof(res);
410 	sc->bge_vpd_readonly = malloc(res.vr_len, M_DEVBUF, M_NOWAIT);
411 	if (sc->bge_vpd_readonly == NULL)
412 		panic("bge_vpd_read");
413 	for (i = 0; i < res.vr_len + 1; i++)
414 		sc->bge_vpd_readonly[i] = bge_vpd_readbyte(sc, i + pos);
415 }
416 #endif
417 
418 /*
419  * Read a byte of data stored in the EEPROM at address 'addr.' The
420  * BCM570x supports both the traditional bitbang interface and an
421  * auto access interface for reading the EEPROM. We use the auto
422  * access method.
423  */
424 static u_int8_t
425 bge_eeprom_getbyte(struct bge_softc *sc, int addr, u_int8_t *dest)
426 {
427 	int i;
428 	u_int32_t byte = 0;
429 
430 	/*
431 	 * Enable use of auto EEPROM access so we can avoid
432 	 * having to use the bitbang method.
433 	 */
434 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_AUTO_EEPROM);
435 
436 	/* Reset the EEPROM, load the clock period. */
437 	CSR_WRITE_4(sc, BGE_EE_ADDR,
438 	    BGE_EEADDR_RESET|BGE_EEHALFCLK(BGE_HALFCLK_384SCL));
439 	DELAY(20);
440 
441 	/* Issue the read EEPROM command. */
442 	CSR_WRITE_4(sc, BGE_EE_ADDR, BGE_EE_READCMD | addr);
443 
444 	/* Wait for completion */
445 	for(i = 0; i < BGE_TIMEOUT * 10; i++) {
446 		DELAY(10);
447 		if (CSR_READ_4(sc, BGE_EE_ADDR) & BGE_EEADDR_DONE)
448 			break;
449 	}
450 
451 	if (i == BGE_TIMEOUT) {
452 		printf("%s: eeprom read timed out\n", sc->bge_dev.dv_xname);
453 		return(0);
454 	}
455 
456 	/* Get result. */
457 	byte = CSR_READ_4(sc, BGE_EE_DATA);
458 
459 	*dest = (byte >> ((addr % 4) * 8)) & 0xFF;
460 
461 	return(0);
462 }
463 
464 /*
465  * Read a sequence of bytes from the EEPROM.
466  */
467 static int
468 bge_read_eeprom(struct bge_softc *sc, caddr_t dest, int off, int cnt)
469 {
470 	int err = 0, i;
471 	u_int8_t byte = 0;
472 
473 	for (i = 0; i < cnt; i++) {
474 		err = bge_eeprom_getbyte(sc, off + i, &byte);
475 		if (err)
476 			break;
477 		*(dest + i) = byte;
478 	}
479 
480 	return(err ? 1 : 0);
481 }
482 
483 static int
484 bge_miibus_readreg(device_t dev, int phy, int reg)
485 {
486 	struct bge_softc *sc = (struct bge_softc *)dev;
487 	u_int32_t val;
488 	u_int32_t saved_autopoll;
489 	int i;
490 
491 	/*
492 	 * Several chips with builtin PHYs will incorrectly answer to
493 	 * other PHY instances than the builtin PHY at id 1.
494 	 */
495 	if (phy != 1 && (sc->bge_quirks & BGE_QUIRK_ONLY_PHY_1))
496 		return(0);
497 
498 	/* Reading with autopolling on may trigger PCI errors */
499 	saved_autopoll = CSR_READ_4(sc, BGE_MI_MODE);
500 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
501 		CSR_WRITE_4(sc, BGE_MI_MODE,
502 		    saved_autopoll &~ BGE_MIMODE_AUTOPOLL);
503 		DELAY(40);
504 	}
505 
506 	CSR_WRITE_4(sc, BGE_MI_COMM, BGE_MICMD_READ|BGE_MICOMM_BUSY|
507 	    BGE_MIPHY(phy)|BGE_MIREG(reg));
508 
509 	for (i = 0; i < BGE_TIMEOUT; i++) {
510 		val = CSR_READ_4(sc, BGE_MI_COMM);
511 		if (!(val & BGE_MICOMM_BUSY))
512 			break;
513 		delay(10);
514 	}
515 
516 	if (i == BGE_TIMEOUT) {
517 		printf("%s: PHY read timed out\n", sc->bge_dev.dv_xname);
518 		val = 0;
519 		goto done;
520 	}
521 
522 	val = CSR_READ_4(sc, BGE_MI_COMM);
523 
524 done:
525 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
526 		CSR_WRITE_4(sc, BGE_MI_MODE, saved_autopoll);
527 		DELAY(40);
528 	}
529 
530 	if (val & BGE_MICOMM_READFAIL)
531 		return(0);
532 
533 	return(val & 0xFFFF);
534 }
535 
536 static void
537 bge_miibus_writereg(device_t dev, int phy, int reg, int val)
538 {
539 	struct bge_softc *sc = (struct bge_softc *)dev;
540 	u_int32_t saved_autopoll;
541 	int i;
542 
543 	/* Touching the PHY while autopolling is on may trigger PCI errors */
544 	saved_autopoll = CSR_READ_4(sc, BGE_MI_MODE);
545 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
546 		delay(40);
547 		CSR_WRITE_4(sc, BGE_MI_MODE,
548 		    saved_autopoll & (~BGE_MIMODE_AUTOPOLL));
549 		delay(10); /* 40 usec is supposed to be adequate */
550 	}
551 
552 	CSR_WRITE_4(sc, BGE_MI_COMM, BGE_MICMD_WRITE|BGE_MICOMM_BUSY|
553 	    BGE_MIPHY(phy)|BGE_MIREG(reg)|val);
554 
555 	for (i = 0; i < BGE_TIMEOUT; i++) {
556 		if (!(CSR_READ_4(sc, BGE_MI_COMM) & BGE_MICOMM_BUSY))
557 			break;
558 		delay(10);
559 	}
560 
561 	if (saved_autopoll & BGE_MIMODE_AUTOPOLL) {
562 		CSR_WRITE_4(sc, BGE_MI_MODE, saved_autopoll);
563 		delay(40);
564 	}
565 
566 	if (i == BGE_TIMEOUT) {
567 		printf("%s: PHY read timed out\n", sc->bge_dev.dv_xname);
568 	}
569 }
570 
571 static void
572 bge_miibus_statchg(device_t dev)
573 {
574 	struct bge_softc *sc = (struct bge_softc *)dev;
575 	struct mii_data *mii = &sc->bge_mii;
576 
577 	/*
578 	 * Get flow control negotiation result.
579 	 */
580 	if (IFM_SUBTYPE(mii->mii_media.ifm_cur->ifm_media) == IFM_AUTO &&
581 	    (mii->mii_media_active & IFM_ETH_FMASK) != sc->bge_flowflags) {
582 		sc->bge_flowflags = mii->mii_media_active & IFM_ETH_FMASK;
583 		mii->mii_media_active &= ~IFM_ETH_FMASK;
584 	}
585 
586 	BGE_CLRBIT(sc, BGE_MAC_MODE, BGE_MACMODE_PORTMODE);
587 	if (IFM_SUBTYPE(mii->mii_media_active) == IFM_1000_T) {
588 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_PORTMODE_GMII);
589 	} else {
590 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_PORTMODE_MII);
591 	}
592 
593 	if ((mii->mii_media_active & IFM_GMASK) == IFM_FDX) {
594 		BGE_CLRBIT(sc, BGE_MAC_MODE, BGE_MACMODE_HALF_DUPLEX);
595 	} else {
596 		BGE_SETBIT(sc, BGE_MAC_MODE, BGE_MACMODE_HALF_DUPLEX);
597 	}
598 
599 	/*
600 	 * 802.3x flow control
601 	 */
602 	if (sc->bge_flowflags & IFM_ETH_RXPAUSE) {
603 		BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_FLOWCTL_ENABLE);
604 	} else {
605 		BGE_CLRBIT(sc, BGE_RX_MODE, BGE_RXMODE_FLOWCTL_ENABLE);
606 	}
607 	if (sc->bge_flowflags & IFM_ETH_TXPAUSE) {
608 		BGE_SETBIT(sc, BGE_TX_MODE, BGE_TXMODE_FLOWCTL_ENABLE);
609 	} else {
610 		BGE_CLRBIT(sc, BGE_TX_MODE, BGE_TXMODE_FLOWCTL_ENABLE);
611 	}
612 }
613 
614 /*
615  * Update rx threshold levels to values in a particular slot
616  * of the interrupt-mitigation table bge_rx_threshes.
617  */
618 static void
619 bge_set_thresh(struct ifnet *ifp, int lvl)
620 {
621 	struct bge_softc *sc = ifp->if_softc;
622 	int s;
623 
624 	/* For now, just save the new Rx-intr thresholds and record
625 	 * that a threshold update is pending.  Updating the hardware
626 	 * registers here (even at splhigh()) is observed to
627 	 * occasionaly cause glitches where Rx-interrupts are not
628 	 * honoured for up to 10 seconds. jonathan@NetBSD.org, 2003-04-05
629 	 */
630 	s = splnet();
631 	sc->bge_rx_coal_ticks = bge_rx_threshes[lvl].rx_ticks;
632 	sc->bge_rx_max_coal_bds = bge_rx_threshes[lvl].rx_max_bds;
633 	sc->bge_pending_rxintr_change = 1;
634 	splx(s);
635 
636 	 return;
637 }
638 
639 
640 /*
641  * Update Rx thresholds of all bge devices
642  */
643 static void
644 bge_update_all_threshes(int lvl)
645 {
646 	struct ifnet *ifp;
647 	const char * const namebuf = "bge";
648 	int namelen;
649 
650 	if (lvl < 0)
651 		lvl = 0;
652 	else if( lvl >= NBGE_RX_THRESH)
653 		lvl = NBGE_RX_THRESH - 1;
654 
655 	namelen = strlen(namebuf);
656 	/*
657 	 * Now search all the interfaces for this name/number
658 	 */
659 	IFNET_FOREACH(ifp) {
660 		if (strncmp(ifp->if_xname, namebuf, namelen) != 0)
661 		      continue;
662 		/* We got a match: update if doing auto-threshold-tuning */
663 		if (bge_auto_thresh)
664 			bge_set_thresh(ifp, lvl);
665 	}
666 }
667 
668 /*
669  * Handle events that have triggered interrupts.
670  */
671 static void
672 bge_handle_events(struct bge_softc *sc)
673 {
674 
675 	return;
676 }
677 
678 /*
679  * Memory management for jumbo frames.
680  */
681 
682 static int
683 bge_alloc_jumbo_mem(struct bge_softc *sc)
684 {
685 	caddr_t			ptr, kva;
686 	bus_dma_segment_t	seg;
687 	int		i, rseg, state, error;
688 	struct bge_jpool_entry   *entry;
689 
690 	state = error = 0;
691 
692 	/* Grab a big chunk o' storage. */
693 	if (bus_dmamem_alloc(sc->bge_dmatag, BGE_JMEM, PAGE_SIZE, 0,
694 	     &seg, 1, &rseg, BUS_DMA_NOWAIT)) {
695 		printf("%s: can't alloc rx buffers\n", sc->bge_dev.dv_xname);
696 		return ENOBUFS;
697 	}
698 
699 	state = 1;
700 	if (bus_dmamem_map(sc->bge_dmatag, &seg, rseg, BGE_JMEM, &kva,
701 	    BUS_DMA_NOWAIT)) {
702 		printf("%s: can't map DMA buffers (%d bytes)\n",
703 		    sc->bge_dev.dv_xname, (int)BGE_JMEM);
704 		error = ENOBUFS;
705 		goto out;
706 	}
707 
708 	state = 2;
709 	if (bus_dmamap_create(sc->bge_dmatag, BGE_JMEM, 1, BGE_JMEM, 0,
710 	    BUS_DMA_NOWAIT, &sc->bge_cdata.bge_rx_jumbo_map)) {
711 		printf("%s: can't create DMA map\n", sc->bge_dev.dv_xname);
712 		error = ENOBUFS;
713 		goto out;
714 	}
715 
716 	state = 3;
717 	if (bus_dmamap_load(sc->bge_dmatag, sc->bge_cdata.bge_rx_jumbo_map,
718 	    kva, BGE_JMEM, NULL, BUS_DMA_NOWAIT)) {
719 		printf("%s: can't load DMA map\n", sc->bge_dev.dv_xname);
720 		error = ENOBUFS;
721 		goto out;
722 	}
723 
724 	state = 4;
725 	sc->bge_cdata.bge_jumbo_buf = (caddr_t)kva;
726 	DPRINTFN(1,("bge_jumbo_buf = %p\n", sc->bge_cdata.bge_jumbo_buf));
727 
728 	SLIST_INIT(&sc->bge_jfree_listhead);
729 	SLIST_INIT(&sc->bge_jinuse_listhead);
730 
731 	/*
732 	 * Now divide it up into 9K pieces and save the addresses
733 	 * in an array.
734 	 */
735 	ptr = sc->bge_cdata.bge_jumbo_buf;
736 	for (i = 0; i < BGE_JSLOTS; i++) {
737 		sc->bge_cdata.bge_jslots[i] = ptr;
738 		ptr += BGE_JLEN;
739 		entry = malloc(sizeof(struct bge_jpool_entry),
740 		    M_DEVBUF, M_NOWAIT);
741 		if (entry == NULL) {
742 			printf("%s: no memory for jumbo buffer queue!\n",
743 			    sc->bge_dev.dv_xname);
744 			error = ENOBUFS;
745 			goto out;
746 		}
747 		entry->slot = i;
748 		SLIST_INSERT_HEAD(&sc->bge_jfree_listhead,
749 				 entry, jpool_entries);
750 	}
751 out:
752 	if (error != 0) {
753 		switch (state) {
754 		case 4:
755 			bus_dmamap_unload(sc->bge_dmatag,
756 			    sc->bge_cdata.bge_rx_jumbo_map);
757 		case 3:
758 			bus_dmamap_destroy(sc->bge_dmatag,
759 			    sc->bge_cdata.bge_rx_jumbo_map);
760 		case 2:
761 			bus_dmamem_unmap(sc->bge_dmatag, kva, BGE_JMEM);
762 		case 1:
763 			bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
764 			break;
765 		default:
766 			break;
767 		}
768 	}
769 
770 	return error;
771 }
772 
773 /*
774  * Allocate a jumbo buffer.
775  */
776 static void *
777 bge_jalloc(struct bge_softc *sc)
778 {
779 	struct bge_jpool_entry   *entry;
780 
781 	entry = SLIST_FIRST(&sc->bge_jfree_listhead);
782 
783 	if (entry == NULL) {
784 		printf("%s: no free jumbo buffers\n", sc->bge_dev.dv_xname);
785 		return(NULL);
786 	}
787 
788 	SLIST_REMOVE_HEAD(&sc->bge_jfree_listhead, jpool_entries);
789 	SLIST_INSERT_HEAD(&sc->bge_jinuse_listhead, entry, jpool_entries);
790 	return(sc->bge_cdata.bge_jslots[entry->slot]);
791 }
792 
793 /*
794  * Release a jumbo buffer.
795  */
796 static void
797 bge_jfree(struct mbuf *m, caddr_t buf, size_t size, void *arg)
798 {
799 	struct bge_jpool_entry *entry;
800 	struct bge_softc *sc;
801 	int i, s;
802 
803 	/* Extract the softc struct pointer. */
804 	sc = (struct bge_softc *)arg;
805 
806 	if (sc == NULL)
807 		panic("bge_jfree: can't find softc pointer!");
808 
809 	/* calculate the slot this buffer belongs to */
810 
811 	i = ((caddr_t)buf
812 	     - (caddr_t)sc->bge_cdata.bge_jumbo_buf) / BGE_JLEN;
813 
814 	if ((i < 0) || (i >= BGE_JSLOTS))
815 		panic("bge_jfree: asked to free buffer that we don't manage!");
816 
817 	s = splvm();
818 	entry = SLIST_FIRST(&sc->bge_jinuse_listhead);
819 	if (entry == NULL)
820 		panic("bge_jfree: buffer not in use!");
821 	entry->slot = i;
822 	SLIST_REMOVE_HEAD(&sc->bge_jinuse_listhead, jpool_entries);
823 	SLIST_INSERT_HEAD(&sc->bge_jfree_listhead, entry, jpool_entries);
824 
825 	if (__predict_true(m != NULL))
826   		pool_cache_put(&mbpool_cache, m);
827 	splx(s);
828 }
829 
830 
831 /*
832  * Intialize a standard receive ring descriptor.
833  */
834 static int
835 bge_newbuf_std(struct bge_softc *sc, int i, struct mbuf *m, bus_dmamap_t dmamap)
836 {
837 	struct mbuf		*m_new = NULL;
838 	struct bge_rx_bd	*r;
839 	int			error;
840 
841 	if (dmamap == NULL) {
842 		error = bus_dmamap_create(sc->bge_dmatag, MCLBYTES, 1,
843 		    MCLBYTES, 0, BUS_DMA_NOWAIT, &dmamap);
844 		if (error != 0)
845 			return error;
846 	}
847 
848 	sc->bge_cdata.bge_rx_std_map[i] = dmamap;
849 
850 	if (m == NULL) {
851 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
852 		if (m_new == NULL) {
853 			return(ENOBUFS);
854 		}
855 
856 		MCLGET(m_new, M_DONTWAIT);
857 		if (!(m_new->m_flags & M_EXT)) {
858 			m_freem(m_new);
859 			return(ENOBUFS);
860 		}
861 		m_new->m_len = m_new->m_pkthdr.len = MCLBYTES;
862 		if (!sc->bge_rx_alignment_bug)
863 		    m_adj(m_new, ETHER_ALIGN);
864 
865 		if (bus_dmamap_load_mbuf(sc->bge_dmatag, dmamap, m_new,
866 		    BUS_DMA_READ|BUS_DMA_NOWAIT))
867 			return(ENOBUFS);
868 	} else {
869 		m_new = m;
870 		m_new->m_len = m_new->m_pkthdr.len = MCLBYTES;
871 		m_new->m_data = m_new->m_ext.ext_buf;
872 		if (!sc->bge_rx_alignment_bug)
873 		    m_adj(m_new, ETHER_ALIGN);
874 	}
875 
876 	sc->bge_cdata.bge_rx_std_chain[i] = m_new;
877 	r = &sc->bge_rdata->bge_rx_std_ring[i];
878 	bge_set_hostaddr(&r->bge_addr,
879 	    dmamap->dm_segs[0].ds_addr);
880 	r->bge_flags = BGE_RXBDFLAG_END;
881 	r->bge_len = m_new->m_len;
882 	r->bge_idx = i;
883 
884 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
885 	    offsetof(struct bge_ring_data, bge_rx_std_ring) +
886 		i * sizeof (struct bge_rx_bd),
887 	    sizeof (struct bge_rx_bd),
888 	    BUS_DMASYNC_PREWRITE|BUS_DMASYNC_PREREAD);
889 
890 	return(0);
891 }
892 
893 /*
894  * Initialize a jumbo receive ring descriptor. This allocates
895  * a jumbo buffer from the pool managed internally by the driver.
896  */
897 static int
898 bge_newbuf_jumbo(struct bge_softc *sc, int i, struct mbuf *m)
899 {
900 	struct mbuf *m_new = NULL;
901 	struct bge_rx_bd *r;
902 
903 	if (m == NULL) {
904 		caddr_t			buf = NULL;
905 
906 		/* Allocate the mbuf. */
907 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
908 		if (m_new == NULL) {
909 			return(ENOBUFS);
910 		}
911 
912 		/* Allocate the jumbo buffer */
913 		buf = bge_jalloc(sc);
914 		if (buf == NULL) {
915 			m_freem(m_new);
916 			printf("%s: jumbo allocation failed "
917 			    "-- packet dropped!\n", sc->bge_dev.dv_xname);
918 			return(ENOBUFS);
919 		}
920 
921 		/* Attach the buffer to the mbuf. */
922 		m_new->m_len = m_new->m_pkthdr.len = BGE_JUMBO_FRAMELEN;
923 		MEXTADD(m_new, buf, BGE_JUMBO_FRAMELEN, M_DEVBUF,
924 		    bge_jfree, sc);
925 		m_new->m_flags |= M_EXT_RW;
926 	} else {
927 		m_new = m;
928 		m_new->m_data = m_new->m_ext.ext_buf;
929 		m_new->m_ext.ext_size = BGE_JUMBO_FRAMELEN;
930 	}
931 
932 	if (!sc->bge_rx_alignment_bug)
933 	    m_adj(m_new, ETHER_ALIGN);
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_HOST_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 			printf("%s: can't alloc txdmamap_pool_entry\n",
1114 			    sc->bge_dev.dv_xname);
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 	struct pci_attach_args	*pa = &(sc->bge_pa);
1219 
1220 
1221 	/* Set endianness before we access any non-PCI registers. */
1222 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL,
1223 	    BGE_INIT);
1224 
1225 	/* Set power state to D0. */
1226 	bge_setpowerstate(sc, 0);
1227 
1228 	/*
1229 	 * Check the 'ROM failed' bit on the RX CPU to see if
1230 	 * self-tests passed.
1231 	 */
1232 	if (CSR_READ_4(sc, BGE_RXCPU_MODE) & BGE_RXCPUMODE_ROMFAIL) {
1233 		printf("%s: RX CPU self-diagnostics failed!\n",
1234 		    sc->bge_dev.dv_xname);
1235 		return(ENODEV);
1236 	}
1237 
1238 	/* Clear the MAC control register */
1239 	CSR_WRITE_4(sc, BGE_MAC_MODE, 0);
1240 
1241 	/*
1242 	 * Clear the MAC statistics block in the NIC's
1243 	 * internal memory.
1244 	 */
1245 	for (i = BGE_STATS_BLOCK;
1246 	    i < BGE_STATS_BLOCK_END + 1; i += sizeof(u_int32_t))
1247 		BGE_MEMWIN_WRITE(pa->pa_pc, pa->pa_tag, i, 0);
1248 
1249 	for (i = BGE_STATUS_BLOCK;
1250 	    i < BGE_STATUS_BLOCK_END + 1; i += sizeof(u_int32_t))
1251 		BGE_MEMWIN_WRITE(pa->pa_pc, pa->pa_tag, i, 0);
1252 
1253 	/* Set up the PCI DMA control register. */
1254 	if (sc->bge_pcie) {
1255 	  u_int32_t device_ctl;
1256 
1257 		/* From FreeBSD */
1258 		DPRINTFN(4, ("(%s: PCI-Express DMA setting)\n",
1259 		    sc->bge_dev.dv_xname));
1260 		dma_rw_ctl = (BGE_PCI_READ_CMD | BGE_PCI_WRITE_CMD |
1261 		    (0xf << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
1262 		    (0x2 << BGE_PCIDMARWCTL_WR_WAT_SHIFT));
1263 
1264 		/* jonathan: alternative from Linux driver */
1265 #define DMA_CTRL_WRITE_PCIE_H20MARK_128         0x00180000
1266 #define DMA_CTRL_WRITE_PCIE_H20MARK_256         0x00380000
1267 
1268 		dma_rw_ctl =   0x76000000; /* XXX XXX XXX */;
1269 		device_ctl = pci_conf_read(pa->pa_pc, pa->pa_tag,
1270 					   BGE_PCI_CONF_DEV_CTRL);
1271 		printf("%s: pcie mode=0x%x\n", sc->bge_dev.dv_xname, device_ctl);
1272 
1273 		if ((device_ctl & 0x00e0) && 0) {
1274 			/*
1275 			 * XXX jonathan@NetBSD.org:
1276 			 * This clause is exactly what the Broadcom-supplied
1277 			 * Linux does; but given overall register programming
1278 			 * by if_bge(4), this larger DMA-write watermark
1279 			 * value causes bcm5721 chips to totally wedge.
1280 			 */
1281 			dma_rw_ctl |= BGE_PCIDMA_RWCTL_PCIE_WRITE_WATRMARK_256;
1282 		} else {
1283 			dma_rw_ctl |= BGE_PCIDMA_RWCTL_PCIE_WRITE_WATRMARK_128;
1284 		}
1285 	} else if (pci_conf_read(pa->pa_pc, pa->pa_tag,BGE_PCI_PCISTATE) &
1286 	    BGE_PCISTATE_PCI_BUSMODE) {
1287 		/* Conventional PCI bus */
1288 	  	DPRINTFN(4, ("(%s: PCI 2.2 DMA setting)\n", sc->bge_dev.dv_xname));
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", sc->bge_dev.dv_xname));
1297 		/* PCI-X bus */
1298 		dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD |
1299 		    (0x3 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
1300 		    (0x3 << BGE_PCIDMARWCTL_WR_WAT_SHIFT) |
1301 		    (0x0F);
1302 		/*
1303 		 * 5703 and 5704 need ONEDMA_AT_ONCE as a workaround
1304 		 * for hardware bugs, which means we should also clear
1305 		 * the low-order MINDMA bits.  In addition, the 5704
1306 		 * uses a different encoding of read/write watermarks.
1307 		 */
1308 		if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5704) {
1309 			dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD |
1310 			  /* should be 0x1f0000 */
1311 			  (0x7 << BGE_PCIDMARWCTL_RD_WAT_SHIFT) |
1312 			  (0x3 << BGE_PCIDMARWCTL_WR_WAT_SHIFT);
1313 			dma_rw_ctl |= BGE_PCIDMARWCTL_ONEDMA_ATONCE;
1314 		}
1315 		else if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5703) {
1316 			dma_rw_ctl &=  0xfffffff0;
1317 			dma_rw_ctl |= BGE_PCIDMARWCTL_ONEDMA_ATONCE;
1318 		}
1319 		else if (BGE_IS_5714_FAMILY(sc)) {
1320 			dma_rw_ctl = BGE_PCI_READ_CMD|BGE_PCI_WRITE_CMD;
1321 			dma_rw_ctl &= ~BGE_PCIDMARWCTL_ONEDMA_ATONCE; /* XXX */
1322 			/* XXX magic values, Broadcom-supplied Linux driver */
1323 			if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5780)
1324 				dma_rw_ctl |= (1 << 20) | (1 << 18) |
1325 				  BGE_PCIDMARWCTL_ONEDMA_ATONCE;
1326 			else
1327 				dma_rw_ctl |= (1<<20) | (1<<18) | (1 << 15);
1328 		}
1329 	}
1330 
1331 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL, dma_rw_ctl);
1332 
1333 	/*
1334 	 * Set up general mode register.
1335 	 */
1336 	CSR_WRITE_4(sc, BGE_MODE_CTL, BGE_DMA_SWAP_OPTIONS|
1337 		    BGE_MODECTL_MAC_ATTN_INTR|BGE_MODECTL_HOST_SEND_BDS|
1338 		    BGE_MODECTL_TX_NO_PHDR_CSUM|BGE_MODECTL_RX_NO_PHDR_CSUM);
1339 
1340 	/* Get cache line size. */
1341 	cachesize = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CACHESZ);
1342 
1343 	/*
1344 	 * Avoid violating PCI spec on certain chip revs.
1345 	 */
1346 	if (pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD) &
1347 	    PCIM_CMD_MWIEN) {
1348 		switch(cachesize) {
1349 		case 1:
1350 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
1351 				   BGE_PCI_WRITE_BNDRY_16BYTES);
1352 			break;
1353 		case 2:
1354 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
1355 				   BGE_PCI_WRITE_BNDRY_32BYTES);
1356 			break;
1357 		case 4:
1358 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
1359 				   BGE_PCI_WRITE_BNDRY_64BYTES);
1360 			break;
1361 		case 8:
1362 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
1363 				   BGE_PCI_WRITE_BNDRY_128BYTES);
1364 			break;
1365 		case 16:
1366 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
1367 				   BGE_PCI_WRITE_BNDRY_256BYTES);
1368 			break;
1369 		case 32:
1370 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
1371 				   BGE_PCI_WRITE_BNDRY_512BYTES);
1372 			break;
1373 		case 64:
1374 			PCI_SETBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_DMA_RW_CTL,
1375 				   BGE_PCI_WRITE_BNDRY_1024BYTES);
1376 			break;
1377 		default:
1378 		/* Disable PCI memory write and invalidate. */
1379 #if 0
1380 			if (bootverbose)
1381 				printf("%s: cache line size %d not "
1382 				    "supported; disabling PCI MWI\n",
1383 				    sc->bge_dev.dv_xname, cachesize);
1384 #endif
1385 			PCI_CLRBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD,
1386 			    PCIM_CMD_MWIEN);
1387 			break;
1388 		}
1389 	}
1390 
1391 	/*
1392 	 * Disable memory write invalidate.  Apparently it is not supported
1393 	 * properly by these devices.
1394 	 */
1395 	PCI_CLRBIT(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD, PCIM_CMD_MWIEN);
1396 
1397 
1398 #ifdef __brokenalpha__
1399 	/*
1400 	 * Must insure that we do not cross an 8K (bytes) boundary
1401 	 * for DMA reads.  Our highest limit is 1K bytes.  This is a
1402 	 * restriction on some ALPHA platforms with early revision
1403 	 * 21174 PCI chipsets, such as the AlphaPC 164lx
1404 	 */
1405 	PCI_SETBIT(sc, BGE_PCI_DMA_RW_CTL, BGE_PCI_READ_BNDRY_1024, 4);
1406 #endif
1407 
1408 	/* Set the timer prescaler (always 66MHz) */
1409 	CSR_WRITE_4(sc, BGE_MISC_CFG, 65 << 1/*BGE_32BITTIME_66MHZ*/);
1410 
1411 	return(0);
1412 }
1413 
1414 static int
1415 bge_blockinit(struct bge_softc *sc)
1416 {
1417 	volatile struct bge_rcb		*rcb;
1418 	bus_size_t		rcb_addr;
1419 	int			i;
1420 	struct ifnet		*ifp = &sc->ethercom.ec_if;
1421 	bge_hostaddr		taddr;
1422 
1423 	/*
1424 	 * Initialize the memory window pointer register so that
1425 	 * we can access the first 32K of internal NIC RAM. This will
1426 	 * allow us to set up the TX send ring RCBs and the RX return
1427 	 * ring RCBs, plus other things which live in NIC memory.
1428 	 */
1429 
1430 	pci_conf_write(sc->bge_pa.pa_pc, sc->bge_pa.pa_tag,
1431 	    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 		printf("%s: buffer manager failed to start\n",
1499 		    sc->bge_dev.dv_xname);
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 		printf("%s: flow-through queue init failed\n",
1516 		    sc->bge_dev.dv_xname);
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 replentish thresholds. The recommended
1587 	 * values are 1/8th the number of descriptors allocated to
1588 	 * each ring.
1589 	 */
1590 	CSR_WRITE_4(sc, BGE_RBDI_STD_REPL_THRESH, BGE_STD_RX_RING_CNT/8);
1591 	CSR_WRITE_4(sc, BGE_RBDI_JUMBO_REPL_THRESH, BGE_JUMBO_RX_RING_CNT/8);
1592 
1593 	/*
1594 	 * Disable all unused send rings by setting the 'ring disabled'
1595 	 * bit in the flags field of all the TX send ring control blocks.
1596 	 * These are located in NIC memory.
1597 	 */
1598 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
1599 	for (i = 0; i < BGE_TX_RINGS_EXTSSRAM_MAX; i++) {
1600 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1601 		    BGE_RCB_MAXLEN_FLAGS(0,BGE_RCB_FLAG_RING_DISABLED));
1602 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
1603 		rcb_addr += sizeof(struct bge_rcb);
1604 	}
1605 
1606 	/* Configure TX RCB 0 (we use only the first ring) */
1607 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
1608 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_tx_ring));
1609 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
1610 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
1611 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr,
1612 		    BGE_NIC_TXRING_ADDR(0, BGE_TX_RING_CNT));
1613 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1614 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1615 		    BGE_RCB_MAXLEN_FLAGS(BGE_TX_RING_CNT, 0));
1616 	}
1617 
1618 	/* Disable all unused RX return rings */
1619 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
1620 	for (i = 0; i < BGE_RX_RINGS_MAX; i++) {
1621 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, 0);
1622 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, 0);
1623 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1624 			    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt,
1625                                      BGE_RCB_FLAG_RING_DISABLED));
1626 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
1627 		CSR_WRITE_4(sc, BGE_MBX_RX_CONS0_LO +
1628 		    (i * (sizeof(u_int64_t))), 0);
1629 		rcb_addr += sizeof(struct bge_rcb);
1630 	}
1631 
1632 	/* Initialize RX ring indexes */
1633 	CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, 0);
1634 	CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, 0);
1635 	CSR_WRITE_4(sc, BGE_MBX_RX_MINI_PROD_LO, 0);
1636 
1637 	/*
1638 	 * Set up RX return ring 0
1639 	 * Note that the NIC address for RX return rings is 0x00000000.
1640 	 * The return rings live entirely within the host, so the
1641 	 * nicaddr field in the RCB isn't used.
1642 	 */
1643 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
1644 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_rx_return_ring));
1645 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
1646 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
1647 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0x00000000);
1648 	RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1649 	    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt, 0));
1650 
1651 	/* Set random backoff seed for TX */
1652 	CSR_WRITE_4(sc, BGE_TX_RANDOM_BACKOFF,
1653 	    LLADDR(ifp->if_sadl)[0] + LLADDR(ifp->if_sadl)[1] +
1654 	    LLADDR(ifp->if_sadl)[2] + LLADDR(ifp->if_sadl)[3] +
1655 	    LLADDR(ifp->if_sadl)[4] + LLADDR(ifp->if_sadl)[5] +
1656 	    BGE_TX_BACKOFF_SEED_MASK);
1657 
1658 	/* Set inter-packet gap */
1659 	CSR_WRITE_4(sc, BGE_TX_LENGTHS, 0x2620);
1660 
1661 	/*
1662 	 * Specify which ring to use for packets that don't match
1663 	 * any RX rules.
1664 	 */
1665 	CSR_WRITE_4(sc, BGE_RX_RULES_CFG, 0x08);
1666 
1667 	/*
1668 	 * Configure number of RX lists. One interrupt distribution
1669 	 * list, sixteen active lists, one bad frames class.
1670 	 */
1671 	CSR_WRITE_4(sc, BGE_RXLP_CFG, 0x181);
1672 
1673 	/* Inialize RX list placement stats mask. */
1674 	CSR_WRITE_4(sc, BGE_RXLP_STATS_ENABLE_MASK, 0x007FFFFF);
1675 	CSR_WRITE_4(sc, BGE_RXLP_STATS_CTL, 0x1);
1676 
1677 	/* Disable host coalescing until we get it set up */
1678 	CSR_WRITE_4(sc, BGE_HCC_MODE, 0x00000000);
1679 
1680 	/* Poll to make sure it's shut down. */
1681 	for (i = 0; i < BGE_TIMEOUT; i++) {
1682 		if (!(CSR_READ_4(sc, BGE_HCC_MODE) & BGE_HCCMODE_ENABLE))
1683 			break;
1684 		DELAY(10);
1685 	}
1686 
1687 	if (i == BGE_TIMEOUT) {
1688 		printf("%s: host coalescing engine failed to idle\n",
1689 		    sc->bge_dev.dv_xname);
1690 		return(ENXIO);
1691 	}
1692 
1693 	/* Set up host coalescing defaults */
1694 	CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, sc->bge_rx_coal_ticks);
1695 	CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS, sc->bge_tx_coal_ticks);
1696 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, sc->bge_rx_max_coal_bds);
1697 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS, sc->bge_tx_max_coal_bds);
1698 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1699 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS_INT, 0);
1700 		CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS_INT, 0);
1701 	}
1702 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS_INT, 0);
1703 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS_INT, 0);
1704 
1705 	/* Set up address of statistics block */
1706 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1707 		bge_set_hostaddr(&taddr,
1708 		    BGE_RING_DMA_ADDR(sc, bge_info.bge_stats));
1709 		CSR_WRITE_4(sc, BGE_HCC_STATS_TICKS, sc->bge_stat_ticks);
1710 		CSR_WRITE_4(sc, BGE_HCC_STATS_BASEADDR, BGE_STATS_BLOCK);
1711 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_HI, taddr.bge_addr_hi);
1712 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_LO, taddr.bge_addr_lo);
1713 	}
1714 
1715 	/* Set up address of status block */
1716 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_status_block));
1717 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_BASEADDR, BGE_STATUS_BLOCK);
1718 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_HI, taddr.bge_addr_hi);
1719 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_LO, taddr.bge_addr_lo);
1720 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx = 0;
1721 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx = 0;
1722 
1723 	/* Turn on host coalescing state machine */
1724 	CSR_WRITE_4(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
1725 
1726 	/* Turn on RX BD completion state machine and enable attentions */
1727 	CSR_WRITE_4(sc, BGE_RBDC_MODE,
1728 	    BGE_RBDCMODE_ENABLE|BGE_RBDCMODE_ATTN);
1729 
1730 	/* Turn on RX list placement state machine */
1731 	CSR_WRITE_4(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
1732 
1733 	/* Turn on RX list selector state machine. */
1734 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1735 		CSR_WRITE_4(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
1736 	}
1737 
1738 	/* Turn on DMA, clear stats */
1739 	CSR_WRITE_4(sc, BGE_MAC_MODE, BGE_MACMODE_TXDMA_ENB|
1740 	    BGE_MACMODE_RXDMA_ENB|BGE_MACMODE_RX_STATS_CLEAR|
1741 	    BGE_MACMODE_TX_STATS_CLEAR|BGE_MACMODE_RX_STATS_ENB|
1742 	    BGE_MACMODE_TX_STATS_ENB|BGE_MACMODE_FRMHDR_DMA_ENB|
1743 	    (sc->bge_tbi ? BGE_PORTMODE_TBI : BGE_PORTMODE_MII));
1744 
1745 	/* Set misc. local control, enable interrupts on attentions */
1746 	sc->bge_local_ctrl_reg = BGE_MLC_INTR_ONATTN | BGE_MLC_AUTO_EEPROM;
1747 
1748 #ifdef notdef
1749 	/* Assert GPIO pins for PHY reset */
1750 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUT0|
1751 	    BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUT2);
1752 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUTEN0|
1753 	    BGE_MLC_MISCIO_OUTEN1|BGE_MLC_MISCIO_OUTEN2);
1754 #endif
1755 
1756 #if defined(not_quite_yet)
1757 	/* Linux driver enables enable gpio pin #1 on 5700s */
1758 	if (sc->bge_chipid == BGE_CHIPID_BCM5700) {
1759 		sc->bge_local_ctrl_reg |=
1760 		  (BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUTEN1);
1761 	}
1762 #endif
1763 	CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
1764 
1765 	/* Turn on DMA completion state machine */
1766 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1767 		CSR_WRITE_4(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
1768 	}
1769 
1770 	/* Turn on write DMA state machine */
1771 	CSR_WRITE_4(sc, BGE_WDMA_MODE,
1772 	    BGE_WDMAMODE_ENABLE|BGE_WDMAMODE_ALL_ATTNS);
1773 
1774 	/* Turn on read DMA state machine */
1775 	{
1776 		uint32_t dma_read_modebits;
1777 
1778 		dma_read_modebits =
1779 		  BGE_RDMAMODE_ENABLE | BGE_RDMAMODE_ALL_ATTNS;
1780 
1781 		if (sc->bge_pcie && 0) {
1782 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_LONG_BURST;
1783 		} else if ((sc->bge_quirks & BGE_QUIRK_5705_CORE)) {
1784 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_SIZE_128;
1785 		}
1786 
1787 		/* XXX broadcom-supplied linux driver; undocumented */
1788 		if (BGE_IS_5750_OR_BEYOND(sc)) {
1789  			/*
1790 			 * XXX: magic values.
1791 			 * From Broadcom-supplied Linux driver;  apparently
1792 			 * required to workaround a DMA bug affecting TSO
1793 			 * on bcm575x/bcm5721?
1794 			 */
1795 			dma_read_modebits |= (1 << 27);
1796 		}
1797 		CSR_WRITE_4(sc, BGE_RDMA_MODE, dma_read_modebits);
1798 	}
1799 
1800 	/* Turn on RX data completion state machine */
1801 	CSR_WRITE_4(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
1802 
1803 	/* Turn on RX BD initiator state machine */
1804 	CSR_WRITE_4(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
1805 
1806 	/* Turn on RX data and RX BD initiator state machine */
1807 	CSR_WRITE_4(sc, BGE_RDBDI_MODE, BGE_RDBDIMODE_ENABLE);
1808 
1809 	/* Turn on Mbuf cluster free state machine */
1810 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1811 		CSR_WRITE_4(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
1812 	}
1813 
1814 	/* Turn on send BD completion state machine */
1815 	CSR_WRITE_4(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
1816 
1817 	/* Turn on send data completion state machine */
1818 	CSR_WRITE_4(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
1819 
1820 	/* Turn on send data initiator state machine */
1821 	if (BGE_IS_5750_OR_BEYOND(sc)) {
1822 		/* XXX: magic value from Linux driver */
1823 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE | 0x08);
1824 	} else {
1825 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
1826 	}
1827 
1828 	/* Turn on send BD initiator state machine */
1829 	CSR_WRITE_4(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
1830 
1831 	/* Turn on send BD selector state machine */
1832 	CSR_WRITE_4(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
1833 
1834 	CSR_WRITE_4(sc, BGE_SDI_STATS_ENABLE_MASK, 0x007FFFFF);
1835 	CSR_WRITE_4(sc, BGE_SDI_STATS_CTL,
1836 	    BGE_SDISTATSCTL_ENABLE|BGE_SDISTATSCTL_FASTER);
1837 
1838 	/* ack/clear link change events */
1839 	CSR_WRITE_4(sc, BGE_MAC_STS, BGE_MACSTAT_SYNC_CHANGED|
1840 	    BGE_MACSTAT_CFG_CHANGED);
1841 	CSR_WRITE_4(sc, BGE_MI_STS, 0);
1842 
1843 	/* Enable PHY auto polling (for MII/GMII only) */
1844 	if (sc->bge_tbi) {
1845 		CSR_WRITE_4(sc, BGE_MI_STS, BGE_MISTS_LINK);
1846  	} else {
1847 		BGE_SETBIT(sc, BGE_MI_MODE, BGE_MIMODE_AUTOPOLL|10<<16);
1848 		if (sc->bge_quirks & BGE_QUIRK_LINK_STATE_BROKEN)
1849 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
1850 			    BGE_EVTENB_MI_INTERRUPT);
1851 	}
1852 
1853 	/* Enable link state change attentions. */
1854 	BGE_SETBIT(sc, BGE_MAC_EVT_ENB, BGE_EVTENB_LINK_CHANGED);
1855 
1856 	return(0);
1857 }
1858 
1859 static const struct bge_revision {
1860 	uint32_t		br_chipid;
1861 	uint32_t		br_quirks;
1862 	const char		*br_name;
1863 } bge_revisions[] = {
1864 	{ BGE_CHIPID_BCM5700_A0,
1865 	  BGE_QUIRK_LINK_STATE_BROKEN,
1866 	  "BCM5700 A0" },
1867 
1868 	{ BGE_CHIPID_BCM5700_A1,
1869 	  BGE_QUIRK_LINK_STATE_BROKEN,
1870 	  "BCM5700 A1" },
1871 
1872 	{ BGE_CHIPID_BCM5700_B0,
1873 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_CSUM_BROKEN|BGE_QUIRK_5700_COMMON,
1874 	  "BCM5700 B0" },
1875 
1876 	{ BGE_CHIPID_BCM5700_B1,
1877 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
1878 	  "BCM5700 B1" },
1879 
1880 	{ BGE_CHIPID_BCM5700_B2,
1881 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
1882 	  "BCM5700 B2" },
1883 
1884 	/* This is treated like a BCM5700 Bx */
1885 	{ BGE_CHIPID_BCM5700_ALTIMA,
1886 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
1887 	  "BCM5700 Altima" },
1888 
1889 	{ BGE_CHIPID_BCM5700_C0,
1890 	  0,
1891 	  "BCM5700 C0" },
1892 
1893 	{ BGE_CHIPID_BCM5701_A0,
1894 	  0, /*XXX really, just not known */
1895 	  "BCM5701 A0" },
1896 
1897 	{ BGE_CHIPID_BCM5701_B0,
1898 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
1899 	  "BCM5701 B0" },
1900 
1901 	{ BGE_CHIPID_BCM5701_B2,
1902 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
1903 	  "BCM5701 B2" },
1904 
1905 	{ BGE_CHIPID_BCM5701_B5,
1906 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
1907 	  "BCM5701 B5" },
1908 
1909 	{ BGE_CHIPID_BCM5703_A0,
1910 	  0,
1911 	  "BCM5703 A0" },
1912 
1913 	{ BGE_CHIPID_BCM5703_A1,
1914 	  0,
1915 	  "BCM5703 A1" },
1916 
1917 	{ BGE_CHIPID_BCM5703_A2,
1918 	  BGE_QUIRK_ONLY_PHY_1,
1919 	  "BCM5703 A2" },
1920 
1921 	{ BGE_CHIPID_BCM5703_A3,
1922 	  BGE_QUIRK_ONLY_PHY_1,
1923 	  "BCM5703 A3" },
1924 
1925 	{ BGE_CHIPID_BCM5704_A0,
1926   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1927 	  "BCM5704 A0" },
1928 
1929 	{ BGE_CHIPID_BCM5704_A1,
1930   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1931 	  "BCM5704 A1" },
1932 
1933 	{ BGE_CHIPID_BCM5704_A2,
1934   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1935 	  "BCM5704 A2" },
1936 
1937 	{ BGE_CHIPID_BCM5704_A3,
1938   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1939 	  "BCM5704 A3" },
1940 
1941 	{ BGE_CHIPID_BCM5705_A0,
1942 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1943 	  "BCM5705 A0" },
1944 
1945 	{ BGE_CHIPID_BCM5705_A1,
1946 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1947 	  "BCM5705 A1" },
1948 
1949 	{ BGE_CHIPID_BCM5705_A2,
1950 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1951 	  "BCM5705 A2" },
1952 
1953 	{ BGE_CHIPID_BCM5705_A3,
1954 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1955 	  "BCM5705 A3" },
1956 
1957 	{ BGE_CHIPID_BCM5750_A0,
1958 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1959 	  "BCM5750 A1" },
1960 
1961 	{ BGE_CHIPID_BCM5750_A1,
1962 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1963 	  "BCM5750 A1" },
1964 
1965 	{ BGE_CHIPID_BCM5751_A1,
1966 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1967 	  "BCM5751 A1" },
1968 
1969 	{ 0, 0, NULL }
1970 };
1971 
1972 /*
1973  * Some defaults for major revisions, so that newer steppings
1974  * that we don't know about have a shot at working.
1975  */
1976 static const struct bge_revision bge_majorrevs[] = {
1977 	{ BGE_ASICREV_BCM5700,
1978 	  BGE_QUIRK_LINK_STATE_BROKEN,
1979 	  "unknown BCM5700" },
1980 
1981 	{ BGE_ASICREV_BCM5701,
1982 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
1983 	  "unknown BCM5701" },
1984 
1985 	{ BGE_ASICREV_BCM5703,
1986 	  0,
1987 	  "unknown BCM5703" },
1988 
1989 	{ BGE_ASICREV_BCM5704,
1990 	  BGE_QUIRK_ONLY_PHY_1,
1991 	  "unknown BCM5704" },
1992 
1993 	{ BGE_ASICREV_BCM5705,
1994 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1995 	  "unknown BCM5705" },
1996 
1997 	{ BGE_ASICREV_BCM5750,
1998 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1999 	  "unknown BCM575x family" },
2000 
2001 	{ BGE_ASICREV_BCM5714,
2002 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2003 	  "unknown BCM5714" },
2004 
2005 	{ BGE_ASICREV_BCM5752,
2006 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2007 	  "unknown BCM5752 family" },
2008 
2009 
2010 	{ BGE_ASICREV_BCM5780,
2011 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2012 	  "unknown BCM5780" },
2013 
2014 	{ BGE_ASICREV_BCM5715,
2015 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2016 	  "unknown BCM5715" },
2017 
2018 	{ 0,
2019 	  0,
2020 	  NULL }
2021 };
2022 
2023 
2024 static const struct bge_revision *
2025 bge_lookup_rev(uint32_t chipid)
2026 {
2027 	const struct bge_revision *br;
2028 
2029 	for (br = bge_revisions; br->br_name != NULL; br++) {
2030 		if (br->br_chipid == chipid)
2031 			return (br);
2032 	}
2033 
2034 	for (br = bge_majorrevs; br->br_name != NULL; br++) {
2035 		if (br->br_chipid == BGE_ASICREV(chipid))
2036 			return (br);
2037 	}
2038 
2039 	return (NULL);
2040 }
2041 
2042 static const struct bge_product {
2043 	pci_vendor_id_t		bp_vendor;
2044 	pci_product_id_t	bp_product;
2045 	const char		*bp_name;
2046 } bge_products[] = {
2047 	/*
2048 	 * The BCM5700 documentation seems to indicate that the hardware
2049 	 * still has the Alteon vendor ID burned into it, though it
2050 	 * should always be overridden by the value in the EEPROM.  We'll
2051 	 * check for it anyway.
2052 	 */
2053 	{ PCI_VENDOR_ALTEON,
2054 	  PCI_PRODUCT_ALTEON_BCM5700,
2055 	  "Broadcom BCM5700 Gigabit Ethernet",
2056 	  },
2057 	{ PCI_VENDOR_ALTEON,
2058 	  PCI_PRODUCT_ALTEON_BCM5701,
2059 	  "Broadcom BCM5701 Gigabit Ethernet",
2060 	  },
2061 
2062 	{ PCI_VENDOR_ALTIMA,
2063 	  PCI_PRODUCT_ALTIMA_AC1000,
2064 	  "Altima AC1000 Gigabit Ethernet",
2065 	  },
2066 	{ PCI_VENDOR_ALTIMA,
2067 	  PCI_PRODUCT_ALTIMA_AC1001,
2068 	  "Altima AC1001 Gigabit Ethernet",
2069 	   },
2070 	{ PCI_VENDOR_ALTIMA,
2071 	  PCI_PRODUCT_ALTIMA_AC9100,
2072 	  "Altima AC9100 Gigabit Ethernet",
2073 	  },
2074 
2075 	{ PCI_VENDOR_BROADCOM,
2076 	  PCI_PRODUCT_BROADCOM_BCM5700,
2077 	  "Broadcom BCM5700 Gigabit Ethernet",
2078 	  },
2079 	{ PCI_VENDOR_BROADCOM,
2080 	  PCI_PRODUCT_BROADCOM_BCM5701,
2081 	  "Broadcom BCM5701 Gigabit Ethernet",
2082 	  },
2083 	{ PCI_VENDOR_BROADCOM,
2084 	  PCI_PRODUCT_BROADCOM_BCM5702,
2085 	  "Broadcom BCM5702 Gigabit Ethernet",
2086 	  },
2087 	{ PCI_VENDOR_BROADCOM,
2088 	  PCI_PRODUCT_BROADCOM_BCM5702X,
2089 	  "Broadcom BCM5702X Gigabit Ethernet" },
2090 
2091 	{ PCI_VENDOR_BROADCOM,
2092 	  PCI_PRODUCT_BROADCOM_BCM5703,
2093 	  "Broadcom BCM5703 Gigabit Ethernet",
2094 	  },
2095 	{ PCI_VENDOR_BROADCOM,
2096 	  PCI_PRODUCT_BROADCOM_BCM5703X,
2097 	  "Broadcom BCM5703X Gigabit Ethernet",
2098 	  },
2099 	{ PCI_VENDOR_BROADCOM,
2100 	  PCI_PRODUCT_BROADCOM_BCM5703A3,
2101 	  "Broadcom BCM5703A3 Gigabit Ethernet",
2102 	  },
2103 
2104    	{ PCI_VENDOR_BROADCOM,
2105 	  PCI_PRODUCT_BROADCOM_BCM5704C,
2106 	  "Broadcom BCM5704C Dual Gigabit Ethernet",
2107 	  },
2108    	{ PCI_VENDOR_BROADCOM,
2109 	  PCI_PRODUCT_BROADCOM_BCM5704S,
2110 	  "Broadcom BCM5704S Dual Gigabit Ethernet",
2111 	  },
2112 
2113    	{ PCI_VENDOR_BROADCOM,
2114 	  PCI_PRODUCT_BROADCOM_BCM5705,
2115 	  "Broadcom BCM5705 Gigabit Ethernet",
2116 	  },
2117    	{ PCI_VENDOR_BROADCOM,
2118 	  PCI_PRODUCT_BROADCOM_BCM5705K,
2119 	  "Broadcom BCM5705K Gigabit Ethernet",
2120 	  },
2121    	{ PCI_VENDOR_BROADCOM,
2122 	  PCI_PRODUCT_BROADCOM_BCM5705_ALT,
2123 	  "Broadcom BCM5705 Gigabit Ethernet",
2124 	  },
2125    	{ PCI_VENDOR_BROADCOM,
2126 	  PCI_PRODUCT_BROADCOM_BCM5705M,
2127 	  "Broadcom BCM5705M Gigabit Ethernet",
2128 	  },
2129 
2130 	{ PCI_VENDOR_BROADCOM,
2131 	  PCI_PRODUCT_BROADCOM_BCM5714,
2132 	  "Broadcom BCM5714/5715 Gigabit Ethernet",
2133 	  },
2134 	{ PCI_VENDOR_BROADCOM,
2135 	  PCI_PRODUCT_BROADCOM_BCM5789,
2136 	  "Broadcom BCM5789 Gigabit Ethernet",
2137 	  },
2138 
2139 	{ PCI_VENDOR_BROADCOM,
2140 	  PCI_PRODUCT_BROADCOM_BCM5721,
2141 	  "Broadcom BCM5721 Gigabit Ethernet",
2142 	  },
2143 
2144 	{ PCI_VENDOR_BROADCOM,
2145 	  PCI_PRODUCT_BROADCOM_BCM5750,
2146 	  "Broadcom BCM5750 Gigabit Ethernet",
2147 	  },
2148 
2149 	{ PCI_VENDOR_BROADCOM,
2150 	  PCI_PRODUCT_BROADCOM_BCM5750M,
2151 	  "Broadcom BCM5750M Gigabit Ethernet",
2152 	  },
2153 
2154 	{ PCI_VENDOR_BROADCOM,
2155 	  PCI_PRODUCT_BROADCOM_BCM5751,
2156 	  "Broadcom BCM5751 Gigabit Ethernet",
2157 	  },
2158 
2159 	{ PCI_VENDOR_BROADCOM,
2160 	  PCI_PRODUCT_BROADCOM_BCM5751M,
2161 	  "Broadcom BCM5751M Gigabit Ethernet",
2162 	  },
2163 
2164 	{ PCI_VENDOR_BROADCOM,
2165 	  PCI_PRODUCT_BROADCOM_BCM5752,
2166 	  "Broadcom BCM5752 Gigabit Ethernet",
2167 	  },
2168 
2169    	{ PCI_VENDOR_BROADCOM,
2170 	  PCI_PRODUCT_BROADCOM_BCM5780,
2171 	  "Broadcom BCM5780 Gigabit Ethernet",
2172 	  },
2173 
2174    	{ PCI_VENDOR_BROADCOM,
2175 	  PCI_PRODUCT_BROADCOM_BCM5780S,
2176 	  "Broadcom BCM5780S Gigabit Ethernet",
2177 	  },
2178 
2179    	{ PCI_VENDOR_BROADCOM,
2180 	  PCI_PRODUCT_BROADCOM_BCM5782,
2181 	  "Broadcom BCM5782 Gigabit Ethernet",
2182 	  },
2183 
2184    	{ PCI_VENDOR_BROADCOM,
2185 	  PCI_PRODUCT_BROADCOM_BCM5788,
2186 	  "Broadcom BCM5788 Gigabit Ethernet",
2187 	  },
2188    	{ PCI_VENDOR_BROADCOM,
2189 	  PCI_PRODUCT_BROADCOM_BCM5789,
2190 	  "Broadcom BCM5789 Gigabit Ethernet",
2191 	  },
2192 
2193    	{ PCI_VENDOR_BROADCOM,
2194 	  PCI_PRODUCT_BROADCOM_BCM5901,
2195 	  "Broadcom BCM5901 Fast Ethernet",
2196 	  },
2197    	{ PCI_VENDOR_BROADCOM,
2198 	  PCI_PRODUCT_BROADCOM_BCM5901A2,
2199 	  "Broadcom BCM5901A2 Fast Ethernet",
2200 	  },
2201 
2202 	{ PCI_VENDOR_SCHNEIDERKOCH,
2203 	  PCI_PRODUCT_SCHNEIDERKOCH_SK_9DX1,
2204 	  "SysKonnect SK-9Dx1 Gigabit Ethernet",
2205 	  },
2206 
2207 	{ PCI_VENDOR_3COM,
2208 	  PCI_PRODUCT_3COM_3C996,
2209 	  "3Com 3c996 Gigabit Ethernet",
2210 	  },
2211 
2212 	{ 0,
2213 	  0,
2214 	  NULL },
2215 };
2216 
2217 static const struct bge_product *
2218 bge_lookup(const struct pci_attach_args *pa)
2219 {
2220 	const struct bge_product *bp;
2221 
2222 	for (bp = bge_products; bp->bp_name != NULL; bp++) {
2223 		if (PCI_VENDOR(pa->pa_id) == bp->bp_vendor &&
2224 		    PCI_PRODUCT(pa->pa_id) == bp->bp_product)
2225 			return (bp);
2226 	}
2227 
2228 	return (NULL);
2229 }
2230 
2231 static int
2232 bge_setpowerstate(struct bge_softc *sc, int powerlevel)
2233 {
2234 #ifdef NOTYET
2235 	u_int32_t pm_ctl = 0;
2236 
2237 	/* XXX FIXME: make sure indirect accesses enabled? */
2238 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_MISC_CTL, 4);
2239 	pm_ctl |= BGE_PCIMISCCTL_INDIRECT_ACCESS;
2240 	pci_write_config(sc->bge_dev, BGE_PCI_MISC_CTL, pm_ctl, 4);
2241 
2242 	/* clear the PME_assert bit and power state bits, enable PME */
2243 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_PWRMGMT_CMD, 2);
2244 	pm_ctl &= ~PCIM_PSTAT_DMASK;
2245 	pm_ctl |= (1 << 8);
2246 
2247 	if (powerlevel == 0) {
2248 		pm_ctl |= PCIM_PSTAT_D0;
2249 		pci_write_config(sc->bge_dev, BGE_PCI_PWRMGMT_CMD,
2250 		    pm_ctl, 2);
2251 		DELAY(10000);
2252 		CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
2253 		DELAY(10000);
2254 
2255 #ifdef NOTYET
2256 		/* XXX FIXME: write 0x02 to phy aux_Ctrl reg */
2257 		bge_miibus_writereg(sc->bge_dev, 1, 0x18, 0x02);
2258 #endif
2259 		DELAY(40); DELAY(40); DELAY(40);
2260 		DELAY(10000);	/* above not quite adequate on 5700 */
2261 		return 0;
2262 	}
2263 
2264 
2265 	/*
2266 	 * Entering ACPI power states D1-D3 is achieved by wiggling
2267 	 * GMII gpio pins. Example code assumes all hardware vendors
2268 	 * followed Broadom's sample pcb layout. Until we verify that
2269 	 * for all supported OEM cards, states D1-D3 are  unsupported.
2270 	 */
2271 	printf("%s: power state %d unimplemented; check GPIO pins\n",
2272 	       sc->bge_dev.dv_xname, powerlevel);
2273 #endif
2274 	return EOPNOTSUPP;
2275 }
2276 
2277 
2278 /*
2279  * Probe for a Broadcom chip. Check the PCI vendor and device IDs
2280  * against our list and return its name if we find a match. Note
2281  * that since the Broadcom controller contains VPD support, we
2282  * can get the device name string from the controller itself instead
2283  * of the compiled-in string. This is a little slow, but it guarantees
2284  * we'll always announce the right product name.
2285  */
2286 static int
2287 bge_probe(device_t parent, cfdata_t match, void *aux)
2288 {
2289 	struct pci_attach_args *pa = (struct pci_attach_args *)aux;
2290 
2291 	if (bge_lookup(pa) != NULL)
2292 		return (1);
2293 
2294 	return (0);
2295 }
2296 
2297 static void
2298 bge_attach(device_t parent, device_t self, void *aux)
2299 {
2300 	struct bge_softc	*sc = (struct bge_softc *)self;
2301 	struct pci_attach_args	*pa = aux;
2302 	const struct bge_product *bp;
2303 	const struct bge_revision *br;
2304 	pci_chipset_tag_t	pc = pa->pa_pc;
2305 	pci_intr_handle_t	ih;
2306 	const char		*intrstr = NULL;
2307 	bus_dma_segment_t	seg;
2308 	int			rseg;
2309 	u_int32_t		hwcfg = 0;
2310 	u_int32_t		mac_addr = 0;
2311 	u_int32_t		command;
2312 	struct ifnet		*ifp;
2313 	caddr_t			kva;
2314 	u_char			eaddr[ETHER_ADDR_LEN];
2315 	pcireg_t		memtype;
2316 	bus_addr_t		memaddr;
2317 	bus_size_t		memsize;
2318 	u_int32_t		pm_ctl;
2319 
2320 	bp = bge_lookup(pa);
2321 	KASSERT(bp != NULL);
2322 
2323 	sc->bge_pa = *pa;
2324 
2325 	aprint_naive(": Ethernet controller\n");
2326 	aprint_normal(": %s\n", bp->bp_name);
2327 
2328 	/*
2329 	 * Map control/status registers.
2330 	 */
2331 	DPRINTFN(5, ("Map control/status regs\n"));
2332 	command = pci_conf_read(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG);
2333 	command |= PCI_COMMAND_MEM_ENABLE | PCI_COMMAND_MASTER_ENABLE;
2334 	pci_conf_write(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG, command);
2335 	command = pci_conf_read(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG);
2336 
2337 	if (!(command & PCI_COMMAND_MEM_ENABLE)) {
2338 		aprint_error("%s: failed to enable memory mapping!\n",
2339 		    sc->bge_dev.dv_xname);
2340 		return;
2341 	}
2342 
2343 	DPRINTFN(5, ("pci_mem_find\n"));
2344 	memtype = pci_mapreg_type(pa->pa_pc, pa->pa_tag, BGE_PCI_BAR0);
2345  	switch (memtype) {
2346 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_32BIT:
2347 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_64BIT:
2348 		if (pci_mapreg_map(pa, BGE_PCI_BAR0,
2349 		    memtype, 0, &sc->bge_btag, &sc->bge_bhandle,
2350 		    &memaddr, &memsize) == 0)
2351 			break;
2352 	default:
2353 		aprint_error("%s: can't find mem space\n",
2354 		    sc->bge_dev.dv_xname);
2355 		return;
2356 	}
2357 
2358 	DPRINTFN(5, ("pci_intr_map\n"));
2359 	if (pci_intr_map(pa, &ih)) {
2360 		aprint_error("%s: couldn't map interrupt\n",
2361 		    sc->bge_dev.dv_xname);
2362 		return;
2363 	}
2364 
2365 	DPRINTFN(5, ("pci_intr_string\n"));
2366 	intrstr = pci_intr_string(pc, ih);
2367 
2368 	DPRINTFN(5, ("pci_intr_establish\n"));
2369 	sc->bge_intrhand = pci_intr_establish(pc, ih, IPL_NET, bge_intr, sc);
2370 
2371 	if (sc->bge_intrhand == NULL) {
2372 		aprint_error("%s: couldn't establish interrupt",
2373 		    sc->bge_dev.dv_xname);
2374 		if (intrstr != NULL)
2375 			aprint_normal(" at %s", intrstr);
2376 		aprint_normal("\n");
2377 		return;
2378 	}
2379 	aprint_normal("%s: interrupting at %s\n",
2380 	    sc->bge_dev.dv_xname, intrstr);
2381 
2382 	/*
2383 	 * Kludge for 5700 Bx bug: a hardware bug (PCIX byte enable?)
2384 	 * can clobber the chip's PCI config-space power control registers,
2385 	 * leaving the card in D3 powersave state.
2386 	 * We do not have memory-mapped registers in this state,
2387 	 * so force device into D0 state before starting initialization.
2388 	 */
2389 	pm_ctl = pci_conf_read(pc, pa->pa_tag, BGE_PCI_PWRMGMT_CMD);
2390 	pm_ctl &= ~(PCI_PWR_D0|PCI_PWR_D1|PCI_PWR_D2|PCI_PWR_D3);
2391 	pm_ctl |= (1 << 8) | PCI_PWR_D0 ; /* D0 state */
2392 	pci_conf_write(pc, pa->pa_tag, BGE_PCI_PWRMGMT_CMD, pm_ctl);
2393 	DELAY(1000);	/* 27 usec is allegedly sufficent */
2394 
2395 	/*
2396 	 * Save ASIC rev.  Look up any quirks associated with this
2397 	 * ASIC.
2398 	 */
2399 	sc->bge_chipid =
2400 	    pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL) &
2401 	    BGE_PCIMISCCTL_ASICREV;
2402 
2403 	/*
2404 	 * Detect PCI-Express devices
2405 	 * XXX: guessed from Linux/FreeBSD; no documentation
2406 	 */
2407 	if (pci_get_capability(pa->pa_pc, pa->pa_tag, PCI_CAP_PCIEXPRESS,
2408 	        NULL, NULL) != 0)
2409 		sc->bge_pcie = 1;
2410 	else
2411 		sc->bge_pcie = 0;
2412 
2413 	/* Try to reset the chip. */
2414 	DPRINTFN(5, ("bge_reset\n"));
2415 	bge_reset(sc);
2416 
2417 	if (bge_chipinit(sc)) {
2418 		aprint_error("%s: chip initialization failed\n",
2419 		    sc->bge_dev.dv_xname);
2420 		bge_release_resources(sc);
2421 		return;
2422 	}
2423 
2424 	/*
2425 	 * Get station address from the EEPROM.
2426 	 */
2427 	mac_addr = bge_readmem_ind(sc, 0x0c14);
2428 	if ((mac_addr >> 16) == 0x484b) {
2429 		eaddr[0] = (u_char)(mac_addr >> 8);
2430 		eaddr[1] = (u_char)(mac_addr >> 0);
2431 		mac_addr = bge_readmem_ind(sc, 0x0c18);
2432 		eaddr[2] = (u_char)(mac_addr >> 24);
2433 		eaddr[3] = (u_char)(mac_addr >> 16);
2434 		eaddr[4] = (u_char)(mac_addr >> 8);
2435 		eaddr[5] = (u_char)(mac_addr >> 0);
2436 	} else if (bge_read_eeprom(sc, (caddr_t)eaddr,
2437 	    BGE_EE_MAC_OFFSET + 2, ETHER_ADDR_LEN)) {
2438 		aprint_error("%s: failed to read station address\n",
2439 		    sc->bge_dev.dv_xname);
2440 		bge_release_resources(sc);
2441 		return;
2442 	}
2443 
2444 	br = bge_lookup_rev(sc->bge_chipid);
2445 	aprint_normal("%s: ", sc->bge_dev.dv_xname);
2446 
2447 	if (br == NULL) {
2448 		aprint_normal("unknown ASIC (0x%04x)", sc->bge_chipid >> 16);
2449 		sc->bge_quirks = 0;
2450 	} else {
2451 		aprint_normal("ASIC %s (0x%04x)",
2452 		    br->br_name, sc->bge_chipid >> 16);
2453 		sc->bge_quirks |= br->br_quirks;
2454 	}
2455 	aprint_normal(", Ethernet address %s\n", ether_sprintf(eaddr));
2456 
2457 	/* Allocate the general information block and ring buffers. */
2458 	if (pci_dma64_available(pa))
2459 		sc->bge_dmatag = pa->pa_dmat64;
2460 	else
2461 		sc->bge_dmatag = pa->pa_dmat;
2462 	DPRINTFN(5, ("bus_dmamem_alloc\n"));
2463 	if (bus_dmamem_alloc(sc->bge_dmatag, sizeof(struct bge_ring_data),
2464 			     PAGE_SIZE, 0, &seg, 1, &rseg, BUS_DMA_NOWAIT)) {
2465 		aprint_error("%s: can't alloc rx buffers\n",
2466 		    sc->bge_dev.dv_xname);
2467 		return;
2468 	}
2469 	DPRINTFN(5, ("bus_dmamem_map\n"));
2470 	if (bus_dmamem_map(sc->bge_dmatag, &seg, rseg,
2471 			   sizeof(struct bge_ring_data), &kva,
2472 			   BUS_DMA_NOWAIT)) {
2473 		aprint_error("%s: can't map DMA buffers (%d bytes)\n",
2474 		    sc->bge_dev.dv_xname, (int)sizeof(struct bge_ring_data));
2475 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2476 		return;
2477 	}
2478 	DPRINTFN(5, ("bus_dmamem_create\n"));
2479 	if (bus_dmamap_create(sc->bge_dmatag, sizeof(struct bge_ring_data), 1,
2480 	    sizeof(struct bge_ring_data), 0,
2481 	    BUS_DMA_NOWAIT, &sc->bge_ring_map)) {
2482 		aprint_error("%s: can't create DMA map\n",
2483 		    sc->bge_dev.dv_xname);
2484 		bus_dmamem_unmap(sc->bge_dmatag, kva,
2485 				 sizeof(struct bge_ring_data));
2486 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2487 		return;
2488 	}
2489 	DPRINTFN(5, ("bus_dmamem_load\n"));
2490 	if (bus_dmamap_load(sc->bge_dmatag, sc->bge_ring_map, kva,
2491 			    sizeof(struct bge_ring_data), NULL,
2492 			    BUS_DMA_NOWAIT)) {
2493 		bus_dmamap_destroy(sc->bge_dmatag, sc->bge_ring_map);
2494 		bus_dmamem_unmap(sc->bge_dmatag, kva,
2495 				 sizeof(struct bge_ring_data));
2496 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2497 		return;
2498 	}
2499 
2500 	DPRINTFN(5, ("bzero\n"));
2501 	sc->bge_rdata = (struct bge_ring_data *)kva;
2502 
2503 	memset(sc->bge_rdata, 0, sizeof(struct bge_ring_data));
2504 
2505 	/* Try to allocate memory for jumbo buffers. */
2506 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
2507 		if (bge_alloc_jumbo_mem(sc)) {
2508 			aprint_error("%s: jumbo buffer allocation failed\n",
2509 			    sc->bge_dev.dv_xname);
2510 		} else
2511 			sc->ethercom.ec_capabilities |= ETHERCAP_JUMBO_MTU;
2512 	}
2513 
2514 	/* Set default tuneable values. */
2515 	sc->bge_stat_ticks = BGE_TICKS_PER_SEC;
2516 	sc->bge_rx_coal_ticks = 150;
2517 	sc->bge_rx_max_coal_bds = 64;
2518 #ifdef ORIG_WPAUL_VALUES
2519 	sc->bge_tx_coal_ticks = 150;
2520 	sc->bge_tx_max_coal_bds = 128;
2521 #else
2522 	sc->bge_tx_coal_ticks = 300;
2523 	sc->bge_tx_max_coal_bds = 400;
2524 #endif
2525 	if (sc->bge_quirks & BGE_QUIRK_5705_CORE) {
2526 		sc->bge_tx_coal_ticks = (12 * 5);
2527 		sc->bge_rx_max_coal_bds = (12 * 5);
2528 			aprint_error("%s: setting short Tx thresholds\n",
2529 			    sc->bge_dev.dv_xname);
2530 	}
2531 
2532 	/* Set up ifnet structure */
2533 	ifp = &sc->ethercom.ec_if;
2534 	ifp->if_softc = sc;
2535 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
2536 	ifp->if_ioctl = bge_ioctl;
2537 	ifp->if_start = bge_start;
2538 	ifp->if_init = bge_init;
2539 	ifp->if_watchdog = bge_watchdog;
2540 	IFQ_SET_MAXLEN(&ifp->if_snd, max(BGE_TX_RING_CNT - 1, IFQ_MAXLEN));
2541 	IFQ_SET_READY(&ifp->if_snd);
2542 	DPRINTFN(5, ("bcopy\n"));
2543 	strcpy(ifp->if_xname, sc->bge_dev.dv_xname);
2544 
2545 	if ((sc->bge_quirks & BGE_QUIRK_CSUM_BROKEN) == 0)
2546 		sc->ethercom.ec_if.if_capabilities |=
2547 		    IFCAP_CSUM_IPv4_Tx | IFCAP_CSUM_IPv4_Rx |
2548 		    IFCAP_CSUM_TCPv4_Tx | IFCAP_CSUM_TCPv4_Rx |
2549 		    IFCAP_CSUM_UDPv4_Tx | IFCAP_CSUM_UDPv4_Rx;
2550 	sc->ethercom.ec_capabilities |=
2551 	    ETHERCAP_VLAN_HWTAGGING | ETHERCAP_VLAN_MTU;
2552 
2553 	if (sc->bge_pcie)
2554 		sc->ethercom.ec_if.if_capabilities |= IFCAP_TSOv4;
2555 
2556 	/*
2557 	 * Do MII setup.
2558 	 */
2559 	DPRINTFN(5, ("mii setup\n"));
2560 	sc->bge_mii.mii_ifp = ifp;
2561 	sc->bge_mii.mii_readreg = bge_miibus_readreg;
2562 	sc->bge_mii.mii_writereg = bge_miibus_writereg;
2563 	sc->bge_mii.mii_statchg = bge_miibus_statchg;
2564 
2565 	/*
2566 	 * Figure out what sort of media we have by checking the
2567 	 * hardware config word in the first 32k of NIC internal memory,
2568 	 * or fall back to the config word in the EEPROM. Note: on some BCM5700
2569 	 * cards, this value appears to be unset. If that's the
2570 	 * case, we have to rely on identifying the NIC by its PCI
2571 	 * subsystem ID, as we do below for the SysKonnect SK-9D41.
2572 	 */
2573 	if (bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_SIG) == BGE_MAGIC_NUMBER) {
2574 		hwcfg = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_NICCFG);
2575 	} else {
2576 		bge_read_eeprom(sc, (caddr_t)&hwcfg,
2577 		    BGE_EE_HWCFG_OFFSET, sizeof(hwcfg));
2578 		hwcfg = be32toh(hwcfg);
2579 	}
2580 	if ((hwcfg & BGE_HWCFG_MEDIA) == BGE_MEDIA_FIBER)
2581 		sc->bge_tbi = 1;
2582 
2583 	/* The SysKonnect SK-9D41 is a 1000baseSX card. */
2584 	if ((pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_SUBSYS) >> 16) ==
2585 	    SK_SUBSYSID_9D41)
2586 		sc->bge_tbi = 1;
2587 
2588 	if (sc->bge_tbi) {
2589 		ifmedia_init(&sc->bge_ifmedia, IFM_IMASK, bge_ifmedia_upd,
2590 		    bge_ifmedia_sts);
2591 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX, 0, NULL);
2592 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX|IFM_FDX,
2593 			    0, NULL);
2594 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO, 0, NULL);
2595 		ifmedia_set(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO);
2596 	} else {
2597 		/*
2598 		 * Do transceiver setup.
2599 		 */
2600 		ifmedia_init(&sc->bge_mii.mii_media, 0, bge_ifmedia_upd,
2601 			     bge_ifmedia_sts);
2602 		mii_attach(&sc->bge_dev, &sc->bge_mii, 0xffffffff,
2603 			   MII_PHY_ANY, MII_OFFSET_ANY,
2604 			   MIIF_FORCEANEG|MIIF_DOPAUSE);
2605 
2606 		if (LIST_FIRST(&sc->bge_mii.mii_phys) == NULL) {
2607 			printf("%s: no PHY found!\n", sc->bge_dev.dv_xname);
2608 			ifmedia_add(&sc->bge_mii.mii_media,
2609 				    IFM_ETHER|IFM_MANUAL, 0, NULL);
2610 			ifmedia_set(&sc->bge_mii.mii_media,
2611 				    IFM_ETHER|IFM_MANUAL);
2612 		} else
2613 			ifmedia_set(&sc->bge_mii.mii_media,
2614 				    IFM_ETHER|IFM_AUTO);
2615 	}
2616 
2617 	/*
2618 	 * When using the BCM5701 in PCI-X mode, data corruption has
2619 	 * been observed in the first few bytes of some received packets.
2620 	 * Aligning the packet buffer in memory eliminates the corruption.
2621 	 * Unfortunately, this misaligns the packet payloads.  On platforms
2622 	 * which do not support unaligned accesses, we will realign the
2623 	 * payloads by copying the received packets.
2624 	 */
2625 	if (sc->bge_quirks & BGE_QUIRK_PCIX_DMA_ALIGN_BUG) {
2626 		/* If in PCI-X mode, work around the alignment bug. */
2627 		if ((pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_PCISTATE) &
2628                     (BGE_PCISTATE_PCI_BUSMODE | BGE_PCISTATE_PCI_BUSSPEED)) ==
2629                          BGE_PCISTATE_PCI_BUSSPEED)
2630 		sc->bge_rx_alignment_bug = 1;
2631         }
2632 
2633 	/*
2634 	 * Call MI attach routine.
2635 	 */
2636 	DPRINTFN(5, ("if_attach\n"));
2637 	if_attach(ifp);
2638 	DPRINTFN(5, ("ether_ifattach\n"));
2639 	ether_ifattach(ifp, eaddr);
2640 #ifdef BGE_EVENT_COUNTERS
2641 	/*
2642 	 * Attach event counters.
2643 	 */
2644 	evcnt_attach_dynamic(&sc->bge_ev_intr, EVCNT_TYPE_INTR,
2645 	    NULL, sc->bge_dev.dv_xname, "intr");
2646 	evcnt_attach_dynamic(&sc->bge_ev_tx_xoff, EVCNT_TYPE_MISC,
2647 	    NULL, sc->bge_dev.dv_xname, "tx_xoff");
2648 	evcnt_attach_dynamic(&sc->bge_ev_tx_xon, EVCNT_TYPE_MISC,
2649 	    NULL, sc->bge_dev.dv_xname, "tx_xon");
2650 	evcnt_attach_dynamic(&sc->bge_ev_rx_xoff, EVCNT_TYPE_MISC,
2651 	    NULL, sc->bge_dev.dv_xname, "rx_xoff");
2652 	evcnt_attach_dynamic(&sc->bge_ev_rx_xon, EVCNT_TYPE_MISC,
2653 	    NULL, sc->bge_dev.dv_xname, "rx_xon");
2654 	evcnt_attach_dynamic(&sc->bge_ev_rx_macctl, EVCNT_TYPE_MISC,
2655 	    NULL, sc->bge_dev.dv_xname, "rx_macctl");
2656 	evcnt_attach_dynamic(&sc->bge_ev_xoffentered, EVCNT_TYPE_MISC,
2657 	    NULL, sc->bge_dev.dv_xname, "xoffentered");
2658 #endif /* BGE_EVENT_COUNTERS */
2659 	DPRINTFN(5, ("callout_init\n"));
2660 	callout_init(&sc->bge_timeout);
2661 
2662 	sc->bge_powerhook = powerhook_establish(bge_powerhook, sc);
2663 	if (sc->bge_powerhook == NULL)
2664 		printf("%s: WARNING: unable to establish PCI power hook\n",
2665 		    sc->bge_dev.dv_xname);
2666 }
2667 
2668 static void
2669 bge_release_resources(struct bge_softc *sc)
2670 {
2671 	if (sc->bge_vpd_prodname != NULL)
2672 		free(sc->bge_vpd_prodname, M_DEVBUF);
2673 
2674 	if (sc->bge_vpd_readonly != NULL)
2675 		free(sc->bge_vpd_readonly, M_DEVBUF);
2676 }
2677 
2678 static void
2679 bge_reset(struct bge_softc *sc)
2680 {
2681 	struct pci_attach_args *pa = &sc->bge_pa;
2682 	u_int32_t cachesize, command, pcistate, new_pcistate;
2683 	int i, val;
2684 
2685 	/* Save some important PCI state. */
2686 	cachesize = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CACHESZ);
2687 	command = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD);
2688 	pcistate = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_PCISTATE);
2689 
2690 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL,
2691 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
2692 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
2693 
2694 	val = BGE_MISCCFG_RESET_CORE_CLOCKS | (65<<1);
2695 	/*
2696 	 * XXX: from FreeBSD/Linux; no documentation
2697 	 */
2698 	if (sc->bge_pcie) {
2699 		if (CSR_READ_4(sc, BGE_PCIE_CTL1) == 0x60)
2700 			CSR_WRITE_4(sc, BGE_PCIE_CTL1, 0x20);
2701 		if (sc->bge_chipid != BGE_CHIPID_BCM5750_A0) {
2702 			/* No idea what that actually means */
2703 			CSR_WRITE_4(sc, BGE_MISC_CFG, 1 << 29);
2704 			val |= (1<<29);
2705 		}
2706 	}
2707 
2708 	/* Issue global reset */
2709 	bge_writereg_ind(sc, BGE_MISC_CFG, val);
2710 
2711 	DELAY(1000);
2712 
2713 	/*
2714 	 * XXX: from FreeBSD/Linux; no documentation
2715 	 */
2716 	if (sc->bge_pcie) {
2717 		if (sc->bge_chipid == BGE_CHIPID_BCM5750_A0) {
2718 			pcireg_t reg;
2719 
2720 			DELAY(500000);
2721 			/* XXX: Magic Numbers */
2722 			reg = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_UNKNOWN0);
2723 			pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_UNKNOWN0,
2724 			    reg | (1 << 15));
2725 		}
2726 		/*
2727 		 * XXX: Magic Numbers.
2728 		 * Sets maximal PCI-e payload and clears any PCI-e errors.
2729 		 * Should be replaced with references to PCI config-space
2730 		 * capability block for PCI-Express.
2731 		 */
2732 		pci_conf_write(pa->pa_pc, pa->pa_tag,
2733 		    BGE_PCI_CONF_DEV_CTRL, 0xf5000);
2734 
2735 	}
2736 
2737 	/* Reset some of the PCI state that got zapped by reset */
2738 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL,
2739 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
2740 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
2741 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD, command);
2742 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_CACHESZ, cachesize);
2743 	bge_writereg_ind(sc, BGE_MISC_CFG, (65 << 1));
2744 
2745 	/* Enable memory arbiter. */
2746 	{
2747 		uint32_t marbmode = 0;
2748 		if (BGE_IS_5714_FAMILY(sc)) {
2749 			marbmode = CSR_READ_4(sc, BGE_MARB_MODE);
2750 		}
2751  		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE | marbmode);
2752 	}
2753 
2754 	/*
2755 	 * Prevent PXE restart: write a magic number to the
2756 	 * general communications memory at 0xB50.
2757 	 */
2758 	bge_writemem_ind(sc, BGE_SOFTWARE_GENCOMM, BGE_MAGIC_NUMBER);
2759 
2760 	/*
2761 	 * Poll the value location we just wrote until
2762 	 * we see the 1's complement of the magic number.
2763 	 * This indicates that the firmware initialization
2764 	 * is complete.
2765 	 */
2766 	for (i = 0; i < BGE_TIMEOUT; i++) {
2767 		val = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM);
2768 		if (val == ~BGE_MAGIC_NUMBER)
2769 			break;
2770 		DELAY(1000);
2771 	}
2772 
2773 	if (i >= BGE_TIMEOUT) {
2774 		printf("%s: firmware handshake timed out, val = %x\n",
2775 		    sc->bge_dev.dv_xname, val);
2776 		/*
2777 		 * XXX: occasionally fired on bcm5721, but without
2778 		 * apparent harm.  For now, keep going if we timeout
2779 		 * against PCI-E devices.
2780 		 */
2781 		 if (!sc->bge_pcie)
2782 		  return;
2783 	}
2784 
2785 	/*
2786 	 * XXX Wait for the value of the PCISTATE register to
2787 	 * return to its original pre-reset state. This is a
2788 	 * fairly good indicator of reset completion. If we don't
2789 	 * wait for the reset to fully complete, trying to read
2790 	 * from the device's non-PCI registers may yield garbage
2791 	 * results.
2792 	 */
2793 	for (i = 0; i < BGE_TIMEOUT; i++) {
2794 		new_pcistate = pci_conf_read(pa->pa_pc, pa->pa_tag,
2795 		    BGE_PCI_PCISTATE);
2796 		if ((new_pcistate & ~BGE_PCISTATE_RESERVED) ==
2797 		    (pcistate & ~BGE_PCISTATE_RESERVED))
2798 			break;
2799 		DELAY(10);
2800 	}
2801 	if ((new_pcistate & ~BGE_PCISTATE_RESERVED) !=
2802 	    (pcistate & ~BGE_PCISTATE_RESERVED)) {
2803 		printf("%s: pcistate failed to revert\n",
2804 		    sc->bge_dev.dv_xname);
2805 	}
2806 
2807 	/* XXX: from FreeBSD/Linux; no documentation */
2808 	if (sc->bge_pcie && sc->bge_chipid != BGE_CHIPID_BCM5750_A0)
2809 		CSR_WRITE_4(sc, BGE_PCIE_CTL0, CSR_READ_4(sc, BGE_PCIE_CTL0) | (1<<25));
2810 
2811 	/* Enable memory arbiter. */
2812 	/* XXX why do this twice? */
2813 	{
2814 		uint32_t marbmode = 0;
2815 		if (BGE_IS_5714_FAMILY(sc)) {
2816 			marbmode = CSR_READ_4(sc, BGE_MARB_MODE);
2817 		}
2818  		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE | marbmode);
2819 	}
2820 
2821 	/* Fix up byte swapping */
2822 	CSR_WRITE_4(sc, BGE_MODE_CTL, BGE_DMA_SWAP_OPTIONS);
2823 
2824 	CSR_WRITE_4(sc, BGE_MAC_MODE, 0);
2825 
2826 	DELAY(10000);
2827 }
2828 
2829 /*
2830  * Frame reception handling. This is called if there's a frame
2831  * on the receive return list.
2832  *
2833  * Note: we have to be able to handle two possibilities here:
2834  * 1) the frame is from the jumbo recieve ring
2835  * 2) the frame is from the standard receive ring
2836  */
2837 
2838 static void
2839 bge_rxeof(struct bge_softc *sc)
2840 {
2841 	struct ifnet *ifp;
2842 	int stdcnt = 0, jumbocnt = 0;
2843 	bus_dmamap_t dmamap;
2844 	bus_addr_t offset, toff;
2845 	bus_size_t tlen;
2846 	int tosync;
2847 
2848 	ifp = &sc->ethercom.ec_if;
2849 
2850 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2851 	    offsetof(struct bge_ring_data, bge_status_block),
2852 	    sizeof (struct bge_status_block),
2853 	    BUS_DMASYNC_POSTREAD);
2854 
2855 	offset = offsetof(struct bge_ring_data, bge_rx_return_ring);
2856 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx -
2857 	    sc->bge_rx_saved_considx;
2858 
2859 	toff = offset + (sc->bge_rx_saved_considx * sizeof (struct bge_rx_bd));
2860 
2861 	if (tosync < 0) {
2862 		tlen = (sc->bge_return_ring_cnt - sc->bge_rx_saved_considx) *
2863 		    sizeof (struct bge_rx_bd);
2864 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2865 		    toff, tlen, BUS_DMASYNC_POSTREAD);
2866 		tosync = -tosync;
2867 	}
2868 
2869 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2870 	    offset, tosync * sizeof (struct bge_rx_bd),
2871 	    BUS_DMASYNC_POSTREAD);
2872 
2873 	while(sc->bge_rx_saved_considx !=
2874 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx) {
2875 		struct bge_rx_bd	*cur_rx;
2876 		u_int32_t		rxidx;
2877 		struct mbuf		*m = NULL;
2878 
2879 		cur_rx = &sc->bge_rdata->
2880 			bge_rx_return_ring[sc->bge_rx_saved_considx];
2881 
2882 		rxidx = cur_rx->bge_idx;
2883 		BGE_INC(sc->bge_rx_saved_considx, sc->bge_return_ring_cnt);
2884 
2885 		if (cur_rx->bge_flags & BGE_RXBDFLAG_JUMBO_RING) {
2886 			BGE_INC(sc->bge_jumbo, BGE_JUMBO_RX_RING_CNT);
2887 			m = sc->bge_cdata.bge_rx_jumbo_chain[rxidx];
2888 			sc->bge_cdata.bge_rx_jumbo_chain[rxidx] = NULL;
2889 			jumbocnt++;
2890 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
2891 				ifp->if_ierrors++;
2892 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
2893 				continue;
2894 			}
2895 			if (bge_newbuf_jumbo(sc, sc->bge_jumbo,
2896 					     NULL)== ENOBUFS) {
2897 				ifp->if_ierrors++;
2898 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
2899 				continue;
2900 			}
2901 		} else {
2902 			BGE_INC(sc->bge_std, BGE_STD_RX_RING_CNT);
2903 			m = sc->bge_cdata.bge_rx_std_chain[rxidx];
2904 			sc->bge_cdata.bge_rx_std_chain[rxidx] = NULL;
2905 			stdcnt++;
2906 			dmamap = sc->bge_cdata.bge_rx_std_map[rxidx];
2907 			sc->bge_cdata.bge_rx_std_map[rxidx] = 0;
2908 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
2909 				ifp->if_ierrors++;
2910 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
2911 				continue;
2912 			}
2913 			if (bge_newbuf_std(sc, sc->bge_std,
2914 			    NULL, dmamap) == ENOBUFS) {
2915 				ifp->if_ierrors++;
2916 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
2917 				continue;
2918 			}
2919 		}
2920 
2921 		ifp->if_ipackets++;
2922 #ifndef __NO_STRICT_ALIGNMENT
2923                 /*
2924                  * XXX: if the 5701 PCIX-Rx-DMA workaround is in effect,
2925                  * the Rx buffer has the layer-2 header unaligned.
2926                  * If our CPU requires alignment, re-align by copying.
2927                  */
2928 		if (sc->bge_rx_alignment_bug) {
2929 			memmove(mtod(m, caddr_t) + ETHER_ALIGN, m->m_data,
2930                                 cur_rx->bge_len);
2931 			m->m_data += ETHER_ALIGN;
2932 		}
2933 #endif
2934 
2935 		m->m_pkthdr.len = m->m_len = cur_rx->bge_len - ETHER_CRC_LEN;
2936 		m->m_pkthdr.rcvif = ifp;
2937 
2938 #if NBPFILTER > 0
2939 		/*
2940 		 * Handle BPF listeners. Let the BPF user see the packet.
2941 		 */
2942 		if (ifp->if_bpf)
2943 			bpf_mtap(ifp->if_bpf, m);
2944 #endif
2945 
2946 		m->m_pkthdr.csum_flags = M_CSUM_IPv4;
2947 
2948 		if ((cur_rx->bge_ip_csum ^ 0xffff) != 0)
2949 			m->m_pkthdr.csum_flags |= M_CSUM_IPv4_BAD;
2950 		/*
2951 		 * Rx transport checksum-offload may also
2952 		 * have bugs with packets which, when transmitted,
2953 		 * were `runts' requiring padding.
2954 		 */
2955 		if (cur_rx->bge_flags & BGE_RXBDFLAG_TCP_UDP_CSUM &&
2956 		    (/* (sc->_bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||*/
2957 		     m->m_pkthdr.len >= ETHER_MIN_NOPAD)) {
2958 			m->m_pkthdr.csum_data =
2959 			    cur_rx->bge_tcp_udp_csum;
2960 			m->m_pkthdr.csum_flags |=
2961 			    (M_CSUM_TCPv4|M_CSUM_UDPv4|
2962 			     M_CSUM_DATA|M_CSUM_NO_PSEUDOHDR);
2963 		}
2964 
2965 		/*
2966 		 * If we received a packet with a vlan tag, pass it
2967 		 * to vlan_input() instead of ether_input().
2968 		 */
2969 		if (cur_rx->bge_flags & BGE_RXBDFLAG_VLAN_TAG)
2970 			VLAN_INPUT_TAG(ifp, m, cur_rx->bge_vlan_tag, continue);
2971 
2972 		(*ifp->if_input)(ifp, m);
2973 	}
2974 
2975 	CSR_WRITE_4(sc, BGE_MBX_RX_CONS0_LO, sc->bge_rx_saved_considx);
2976 	if (stdcnt)
2977 		CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, sc->bge_std);
2978 	if (jumbocnt)
2979 		CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, sc->bge_jumbo);
2980 }
2981 
2982 static void
2983 bge_txeof(struct bge_softc *sc)
2984 {
2985 	struct bge_tx_bd *cur_tx = NULL;
2986 	struct ifnet *ifp;
2987 	struct txdmamap_pool_entry *dma;
2988 	bus_addr_t offset, toff;
2989 	bus_size_t tlen;
2990 	int tosync;
2991 	struct mbuf *m;
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_tx_ring);
3001 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx -
3002 	    sc->bge_tx_saved_considx;
3003 
3004 	toff = offset + (sc->bge_tx_saved_considx * sizeof (struct bge_tx_bd));
3005 
3006 	if (tosync < 0) {
3007 		tlen = (BGE_TX_RING_CNT - sc->bge_tx_saved_considx) *
3008 		    sizeof (struct bge_tx_bd);
3009 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
3010 		    toff, tlen, BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
3011 		tosync = -tosync;
3012 	}
3013 
3014 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
3015 	    offset, tosync * sizeof (struct bge_tx_bd),
3016 	    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
3017 
3018 	/*
3019 	 * Go through our tx ring and free mbufs for those
3020 	 * frames that have been sent.
3021 	 */
3022 	while (sc->bge_tx_saved_considx !=
3023 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx) {
3024 		u_int32_t		idx = 0;
3025 
3026 		idx = sc->bge_tx_saved_considx;
3027 		cur_tx = &sc->bge_rdata->bge_tx_ring[idx];
3028 		if (cur_tx->bge_flags & BGE_TXBDFLAG_END)
3029 			ifp->if_opackets++;
3030 		m = sc->bge_cdata.bge_tx_chain[idx];
3031 		if (m != NULL) {
3032 			sc->bge_cdata.bge_tx_chain[idx] = NULL;
3033 			dma = sc->txdma[idx];
3034 			bus_dmamap_sync(sc->bge_dmatag, dma->dmamap, 0,
3035 			    dma->dmamap->dm_mapsize, BUS_DMASYNC_POSTWRITE);
3036 			bus_dmamap_unload(sc->bge_dmatag, dma->dmamap);
3037 			SLIST_INSERT_HEAD(&sc->txdma_list, dma, link);
3038 			sc->txdma[idx] = NULL;
3039 
3040 			m_freem(m);
3041 		}
3042 		sc->bge_txcnt--;
3043 		BGE_INC(sc->bge_tx_saved_considx, BGE_TX_RING_CNT);
3044 		ifp->if_timer = 0;
3045 	}
3046 
3047 	if (cur_tx != NULL)
3048 		ifp->if_flags &= ~IFF_OACTIVE;
3049 }
3050 
3051 static int
3052 bge_intr(void *xsc)
3053 {
3054 	struct bge_softc *sc;
3055 	struct ifnet *ifp;
3056 
3057 	sc = xsc;
3058 	ifp = &sc->ethercom.ec_if;
3059 
3060 #ifdef notdef
3061 	/* Avoid this for now -- checking this register is expensive. */
3062 	/* Make sure this is really our interrupt. */
3063 	if (!(CSR_READ_4(sc, BGE_MISC_LOCAL_CTL) & BGE_MLC_INTR_STATE))
3064 		return (0);
3065 #endif
3066 	/* Ack interrupt and stop others from occuring. */
3067 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 1);
3068 
3069 	BGE_EVCNT_INCR(sc->bge_ev_intr);
3070 
3071 	/*
3072 	 * Process link state changes.
3073 	 * Grrr. The link status word in the status block does
3074 	 * not work correctly on the BCM5700 rev AX and BX chips,
3075 	 * according to all available information. Hence, we have
3076 	 * to enable MII interrupts in order to properly obtain
3077 	 * async link changes. Unfortunately, this also means that
3078 	 * we have to read the MAC status register to detect link
3079 	 * changes, thereby adding an additional register access to
3080 	 * the interrupt handler.
3081 	 */
3082 
3083 	if (sc->bge_quirks & BGE_QUIRK_LINK_STATE_BROKEN) {
3084 		u_int32_t		status;
3085 
3086 		status = CSR_READ_4(sc, BGE_MAC_STS);
3087 		if (status & BGE_MACSTAT_MI_INTERRUPT) {
3088 			sc->bge_link = 0;
3089 			callout_stop(&sc->bge_timeout);
3090 			bge_tick(sc);
3091 			/* Clear the interrupt */
3092 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
3093 			    BGE_EVTENB_MI_INTERRUPT);
3094 			bge_miibus_readreg(&sc->bge_dev, 1, BRGPHY_MII_ISR);
3095 			bge_miibus_writereg(&sc->bge_dev, 1, BRGPHY_MII_IMR,
3096 			    BRGPHY_INTRS);
3097 		}
3098 	} else {
3099 		if (sc->bge_rdata->bge_status_block.bge_status &
3100 		    BGE_STATFLAG_LINKSTATE_CHANGED) {
3101 			sc->bge_link = 0;
3102 			callout_stop(&sc->bge_timeout);
3103 			bge_tick(sc);
3104 			/* Clear the interrupt */
3105 			CSR_WRITE_4(sc, BGE_MAC_STS, BGE_MACSTAT_SYNC_CHANGED|
3106 			    BGE_MACSTAT_CFG_CHANGED|BGE_MACSTAT_MI_COMPLETE|
3107 			    BGE_MACSTAT_LINK_CHANGED);
3108 		}
3109 	}
3110 
3111 	if (ifp->if_flags & IFF_RUNNING) {
3112 		/* Check RX return ring producer/consumer */
3113 		bge_rxeof(sc);
3114 
3115 		/* Check TX ring producer/consumer */
3116 		bge_txeof(sc);
3117 	}
3118 
3119 	if (sc->bge_pending_rxintr_change) {
3120 		uint32_t rx_ticks = sc->bge_rx_coal_ticks;
3121 		uint32_t rx_bds = sc->bge_rx_max_coal_bds;
3122 		uint32_t junk;
3123 
3124 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, rx_ticks);
3125 		DELAY(10);
3126 		junk = CSR_READ_4(sc, BGE_HCC_RX_COAL_TICKS);
3127 
3128 		CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, rx_bds);
3129 		DELAY(10);
3130 		junk = CSR_READ_4(sc, BGE_HCC_RX_MAX_COAL_BDS);
3131 
3132 		sc->bge_pending_rxintr_change = 0;
3133 	}
3134 	bge_handle_events(sc);
3135 
3136 	/* Re-enable interrupts. */
3137 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 0);
3138 
3139 	if (ifp->if_flags & IFF_RUNNING && !IFQ_IS_EMPTY(&ifp->if_snd))
3140 		bge_start(ifp);
3141 
3142 	return (1);
3143 }
3144 
3145 static void
3146 bge_tick(void *xsc)
3147 {
3148 	struct bge_softc *sc = xsc;
3149 	struct mii_data *mii = &sc->bge_mii;
3150 	struct ifmedia *ifm = NULL;
3151 	struct ifnet *ifp = &sc->ethercom.ec_if;
3152 	int s;
3153 
3154 	s = splnet();
3155 
3156 	bge_stats_update(sc);
3157 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
3158 	if (sc->bge_link) {
3159 		splx(s);
3160 		return;
3161 	}
3162 
3163 	if (sc->bge_tbi) {
3164 		ifm = &sc->bge_ifmedia;
3165 		if (CSR_READ_4(sc, BGE_MAC_STS) &
3166 		    BGE_MACSTAT_TBI_PCS_SYNCHED) {
3167 			sc->bge_link++;
3168 			CSR_WRITE_4(sc, BGE_MAC_STS, 0xFFFFFFFF);
3169 			if (!IFQ_IS_EMPTY(&ifp->if_snd))
3170 				bge_start(ifp);
3171 		}
3172 		splx(s);
3173 		return;
3174 	}
3175 
3176 	mii_tick(mii);
3177 
3178 	if (!sc->bge_link && mii->mii_media_status & IFM_ACTIVE &&
3179 	    IFM_SUBTYPE(mii->mii_media_active) != IFM_NONE) {
3180 		sc->bge_link++;
3181 		if (!IFQ_IS_EMPTY(&ifp->if_snd))
3182 			bge_start(ifp);
3183 	}
3184 
3185 	splx(s);
3186 }
3187 
3188 static void
3189 bge_stats_update(struct bge_softc *sc)
3190 {
3191 	struct ifnet *ifp = &sc->ethercom.ec_if;
3192 	bus_size_t stats = BGE_MEMWIN_START + BGE_STATS_BLOCK;
3193 	bus_size_t rstats = BGE_RX_STATS;
3194 
3195 #define READ_RSTAT(sc, stats, stat) \
3196 	  CSR_READ_4(sc, stats + offsetof(struct bge_mac_stats_regs, stat))
3197 
3198 	if (sc->bge_quirks & BGE_QUIRK_5705_CORE) {
3199 		ifp->if_collisions +=
3200 		    READ_RSTAT(sc, rstats, dot3StatsSingleCollisionFrames) +
3201 		    READ_RSTAT(sc, rstats, dot3StatsMultipleCollisionFrames) +
3202 		    READ_RSTAT(sc, rstats, dot3StatsExcessiveCollisions) +
3203 		    READ_RSTAT(sc, rstats, dot3StatsLateCollisions);
3204 
3205 		BGE_EVCNT_ADD(sc->bge_ev_tx_xoff,
3206 			      READ_RSTAT(sc, rstats, outXoffSent));
3207 		BGE_EVCNT_ADD(sc->bge_ev_tx_xon,
3208 			      READ_RSTAT(sc, rstats, outXonSent));
3209 		BGE_EVCNT_ADD(sc->bge_ev_rx_xoff,
3210 			      READ_RSTAT(sc, rstats, xoffPauseFramesReceived));
3211 		BGE_EVCNT_ADD(sc->bge_ev_rx_xon,
3212 			      READ_RSTAT(sc, rstats, xonPauseFramesReceived));
3213 		BGE_EVCNT_ADD(sc->bge_ev_rx_macctl,
3214 			      READ_RSTAT(sc, rstats, macControlFramesReceived));
3215 		BGE_EVCNT_ADD(sc->bge_ev_xoffentered,
3216 			      READ_RSTAT(sc, rstats, xoffStateEntered));
3217 		return;
3218 	}
3219 
3220 #undef READ_RSTAT
3221 #define READ_STAT(sc, stats, stat) \
3222 	  CSR_READ_4(sc, stats + offsetof(struct bge_stats, stat))
3223 
3224 	ifp->if_collisions +=
3225 	  (READ_STAT(sc, stats, dot3StatsSingleCollisionFrames.bge_addr_lo) +
3226 	   READ_STAT(sc, stats, dot3StatsMultipleCollisionFrames.bge_addr_lo) +
3227 	   READ_STAT(sc, stats, dot3StatsExcessiveCollisions.bge_addr_lo) +
3228 	   READ_STAT(sc, stats, dot3StatsLateCollisions.bge_addr_lo)) -
3229 	  ifp->if_collisions;
3230 
3231 	BGE_EVCNT_UPD(sc->bge_ev_tx_xoff,
3232 		      READ_STAT(sc, stats, outXoffSent.bge_addr_lo));
3233 	BGE_EVCNT_UPD(sc->bge_ev_tx_xon,
3234 		      READ_STAT(sc, stats, outXonSent.bge_addr_lo));
3235 	BGE_EVCNT_UPD(sc->bge_ev_rx_xoff,
3236 		      READ_STAT(sc, stats,
3237 		      		xoffPauseFramesReceived.bge_addr_lo));
3238 	BGE_EVCNT_UPD(sc->bge_ev_rx_xon,
3239 		      READ_STAT(sc, stats, xonPauseFramesReceived.bge_addr_lo));
3240 	BGE_EVCNT_UPD(sc->bge_ev_rx_macctl,
3241 		      READ_STAT(sc, stats,
3242 		      		macControlFramesReceived.bge_addr_lo));
3243 	BGE_EVCNT_UPD(sc->bge_ev_xoffentered,
3244 		      READ_STAT(sc, stats, xoffStateEntered.bge_addr_lo));
3245 
3246 #undef READ_STAT
3247 
3248 #ifdef notdef
3249 	ifp->if_collisions +=
3250 	   (sc->bge_rdata->bge_info.bge_stats.dot3StatsSingleCollisionFrames +
3251 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsMultipleCollisionFrames +
3252 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsExcessiveCollisions +
3253 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsLateCollisions) -
3254 	   ifp->if_collisions;
3255 #endif
3256 }
3257 
3258 /*
3259  * Pad outbound frame to ETHER_MIN_NOPAD for an unusual reason.
3260  * The bge hardware will pad out Tx runts to ETHER_MIN_NOPAD,
3261  * but when such padded frames employ the  bge IP/TCP checksum offload,
3262  * the hardware checksum assist gives incorrect results (possibly
3263  * from incorporating its own padding into the UDP/TCP checksum; who knows).
3264  * If we pad such runts with zeros, the onboard checksum comes out correct.
3265  */
3266 static inline int
3267 bge_cksum_pad(struct mbuf *pkt)
3268 {
3269 	struct mbuf *last = NULL;
3270 	int padlen;
3271 
3272 	padlen = ETHER_MIN_NOPAD - pkt->m_pkthdr.len;
3273 
3274 	/* if there's only the packet-header and we can pad there, use it. */
3275 	if (pkt->m_pkthdr.len == pkt->m_len &&
3276 	    !M_READONLY(pkt) && M_TRAILINGSPACE(pkt) >= padlen) {
3277 		last = pkt;
3278 	} else {
3279 		/*
3280 		 * Walk packet chain to find last mbuf. We will either
3281 		 * pad there, or append a new mbuf and pad it
3282 		 * (thus perhaps avoiding the bcm5700 dma-min bug).
3283 		 */
3284 		for (last = pkt; last->m_next != NULL; last = last->m_next) {
3285 	      	       (void) 0; /* do nothing*/
3286 		}
3287 
3288 		/* `last' now points to last in chain. */
3289 		if (!M_READONLY(last) && M_TRAILINGSPACE(last) >= padlen) {
3290 			(void) 0; /* we can pad here, in-place. */
3291 		} else {
3292 			/* Allocate new empty mbuf, pad it. Compact later. */
3293 			struct mbuf *n;
3294 			MGET(n, M_DONTWAIT, MT_DATA);
3295 			n->m_len = 0;
3296 			last->m_next = n;
3297 			last = n;
3298 		}
3299 	}
3300 
3301 #ifdef DEBUG
3302 	  /*KASSERT(M_WRITABLE(last), ("to-pad mbuf not writeable\n"));*/
3303 	  KASSERT(M_TRAILINGSPACE(last) >= padlen /*, ("insufficient space to pad\n")*/ );
3304 #endif
3305 	/* Now zero the pad area, to avoid the bge cksum-assist bug */
3306 	memset(mtod(last, caddr_t) + last->m_len, 0, padlen);
3307 	last->m_len += padlen;
3308 	pkt->m_pkthdr.len += padlen;
3309 	return 0;
3310 }
3311 
3312 /*
3313  * Compact outbound packets to avoid bug with DMA segments less than 8 bytes.
3314  */
3315 static inline int
3316 bge_compact_dma_runt(struct mbuf *pkt)
3317 {
3318 	struct mbuf	*m, *prev;
3319 	int 		totlen, prevlen;
3320 
3321 	prev = NULL;
3322 	totlen = 0;
3323 	prevlen = -1;
3324 
3325 	for (m = pkt; m != NULL; prev = m,m = m->m_next) {
3326 		int mlen = m->m_len;
3327 		int shortfall = 8 - mlen ;
3328 
3329 		totlen += mlen;
3330 		if (mlen == 0) {
3331 			continue;
3332 		}
3333 		if (mlen >= 8)
3334 			continue;
3335 
3336 		/* If we get here, mbuf data is too small for DMA engine.
3337 		 * Try to fix by shuffling data to prev or next in chain.
3338 		 * If that fails, do a compacting deep-copy of the whole chain.
3339 		 */
3340 
3341 		/* Internal frag. If fits in prev, copy it there. */
3342 		if (prev && !M_READONLY(prev) &&
3343 		      M_TRAILINGSPACE(prev) >= m->m_len) {
3344 		  	bcopy(m->m_data,
3345 			      prev->m_data+prev->m_len,
3346 			      mlen);
3347 			prev->m_len += mlen;
3348 			m->m_len = 0;
3349 			/* XXX stitch chain */
3350 			prev->m_next = m_free(m);
3351 			m = prev;
3352 			continue;
3353 		}
3354 		else if (m->m_next != NULL && !M_READONLY(m) &&
3355 			     M_TRAILINGSPACE(m) >= shortfall &&
3356 			     m->m_next->m_len >= (8 + shortfall)) {
3357 		    /* m is writable and have enough data in next, pull up. */
3358 
3359 		  	bcopy(m->m_next->m_data,
3360 			      m->m_data+m->m_len,
3361 			      shortfall);
3362 			m->m_len += shortfall;
3363 			m->m_next->m_len -= shortfall;
3364 			m->m_next->m_data += shortfall;
3365 		}
3366 		else if (m->m_next == NULL || 1) {
3367 		  	/* Got a runt at the very end of the packet.
3368 			 * borrow data from the tail of the preceding mbuf and
3369 			 * update its length in-place. (The original data is still
3370 			 * valid, so we can do this even if prev is not writable.)
3371 			 */
3372 
3373 			/* if we'd make prev a runt, just move all of its data. */
3374 #ifdef DEBUG
3375 			KASSERT(prev != NULL /*, ("runt but null PREV")*/);
3376 			KASSERT(prev->m_len >= 8 /*, ("runt prev")*/);
3377 #endif
3378 			if ((prev->m_len - shortfall) < 8)
3379 				shortfall = prev->m_len;
3380 
3381 #ifdef notyet	/* just do the safe slow thing for now */
3382 			if (!M_READONLY(m)) {
3383 				if (M_LEADINGSPACE(m) < shorfall) {
3384 					void *m_dat;
3385 					m_dat = (m->m_flags & M_PKTHDR) ?
3386 					  m->m_pktdat : m->dat;
3387 					memmove(m_dat, mtod(m, void*), m->m_len);
3388 					m->m_data = m_dat;
3389 				    }
3390 			} else
3391 #endif	/* just do the safe slow thing */
3392 			{
3393 				struct mbuf * n = NULL;
3394 				int newprevlen = prev->m_len - shortfall;
3395 
3396 				MGET(n, M_NOWAIT, MT_DATA);
3397 				if (n == NULL)
3398 				   return ENOBUFS;
3399 				KASSERT(m->m_len + shortfall < MLEN
3400 					/*,
3401 					  ("runt %d +prev %d too big\n", m->m_len, shortfall)*/);
3402 
3403 				/* first copy the data we're stealing from prev */
3404 				bcopy(prev->m_data + newprevlen, n->m_data, shortfall);
3405 
3406 				/* update prev->m_len accordingly */
3407 				prev->m_len -= shortfall;
3408 
3409 				/* copy data from runt m */
3410 				bcopy(m->m_data, n->m_data + shortfall, m->m_len);
3411 
3412 				/* n holds what we stole from prev, plus m */
3413 				n->m_len = shortfall + m->m_len;
3414 
3415 				/* stitch n into chain and free m */
3416 				n->m_next = m->m_next;
3417 				prev->m_next = n;
3418 				/* KASSERT(m->m_next == NULL); */
3419 				m->m_next = NULL;
3420 				m_free(m);
3421 				m = n;	/* for continuing loop */
3422 			}
3423 		}
3424 		prevlen = m->m_len;
3425 	}
3426 	return 0;
3427 }
3428 
3429 /*
3430  * Encapsulate an mbuf chain in the tx ring  by coupling the mbuf data
3431  * pointers to descriptors.
3432  */
3433 static int
3434 bge_encap(struct bge_softc *sc, struct mbuf *m_head, u_int32_t *txidx)
3435 {
3436 	struct bge_tx_bd	*f = NULL;
3437 	u_int32_t		frag, cur, cnt = 0;
3438 	u_int16_t		csum_flags = 0;
3439 	u_int16_t		txbd_tso_flags = 0;
3440 	struct txdmamap_pool_entry *dma;
3441 	bus_dmamap_t dmamap;
3442 	int			i = 0;
3443 	struct m_tag		*mtag;
3444 	int			use_tso, maxsegsize, error;
3445 
3446 	cur = frag = *txidx;
3447 
3448 	if (m_head->m_pkthdr.csum_flags) {
3449 		if (m_head->m_pkthdr.csum_flags & M_CSUM_IPv4)
3450 			csum_flags |= BGE_TXBDFLAG_IP_CSUM;
3451 		if (m_head->m_pkthdr.csum_flags & (M_CSUM_TCPv4|M_CSUM_UDPv4))
3452 			csum_flags |= BGE_TXBDFLAG_TCP_UDP_CSUM;
3453 	}
3454 
3455 	/*
3456 	 * If we were asked to do an outboard checksum, and the NIC
3457 	 * has the bug where it sometimes adds in the Ethernet padding,
3458 	 * explicitly pad with zeros so the cksum will be correct either way.
3459 	 * (For now, do this for all chip versions, until newer
3460 	 * are confirmed to not require the workaround.)
3461 	 */
3462 	if ((csum_flags & BGE_TXBDFLAG_TCP_UDP_CSUM) == 0 ||
3463 #ifdef notyet
3464 	    (sc->bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||
3465 #endif
3466 	    m_head->m_pkthdr.len >= ETHER_MIN_NOPAD)
3467 		goto check_dma_bug;
3468 
3469 	if (bge_cksum_pad(m_head) != 0) {
3470 	    return ENOBUFS;
3471 	}
3472 
3473 check_dma_bug:
3474 	if (!(sc->bge_quirks & BGE_QUIRK_5700_SMALLDMA))
3475 		goto doit;
3476 	/*
3477 	 * bcm5700 Revision B silicon cannot handle DMA descriptors with
3478 	 * less than eight bytes.  If we encounter a teeny mbuf
3479 	 * at the end of a chain, we can pad.  Otherwise, copy.
3480 	 */
3481 	if (bge_compact_dma_runt(m_head) != 0)
3482 		return ENOBUFS;
3483 
3484 doit:
3485 	dma = SLIST_FIRST(&sc->txdma_list);
3486 	if (dma == NULL)
3487 		return ENOBUFS;
3488 	dmamap = dma->dmamap;
3489 
3490 	/*
3491 	 * Set up any necessary TSO state before we start packing...
3492 	 */
3493 	use_tso = (m_head->m_pkthdr.csum_flags & M_CSUM_TSOv4) != 0;
3494 	if (!use_tso) {
3495 		maxsegsize = 0;
3496 	} else {	/* TSO setup */
3497 		unsigned  mss;
3498 		struct ether_header *eh;
3499 		unsigned ip_tcp_hlen, iptcp_opt_words, tcp_seg_flags, offset;
3500 		struct mbuf * m0 = m_head;
3501 		struct ip *ip;
3502 		struct tcphdr *th;
3503 		int iphl, hlen;
3504 
3505 		/*
3506 		 * XXX It would be nice if the mbuf pkthdr had offset
3507 		 * fields for the protocol headers.
3508 		 */
3509 
3510 		eh = mtod(m0, struct ether_header *);
3511 		switch (htons(eh->ether_type)) {
3512 		case ETHERTYPE_IP:
3513 			offset = ETHER_HDR_LEN;
3514 			break;
3515 
3516 		case ETHERTYPE_VLAN:
3517 			offset = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3518 			break;
3519 
3520 		default:
3521 			/*
3522 			 * Don't support this protocol or encapsulation.
3523 			 */
3524 			return (ENOBUFS);
3525 		}
3526 
3527 		/*
3528 		 * TCP/IP headers are in the first mbuf; we can do
3529 		 * this the easy way.
3530 		 */
3531 		iphl = M_CSUM_DATA_IPv4_IPHL(m0->m_pkthdr.csum_data);
3532 		hlen = iphl + offset;
3533 		if (__predict_false(m0->m_len <
3534 				    (hlen + sizeof(struct tcphdr)))) {
3535 
3536 			  printf("TSO: hard case m0->m_len == %d <"
3537 				 " ip/tcp hlen %zd, not handled yet\n",
3538 				 m0->m_len, hlen+ sizeof(struct tcphdr));
3539 #ifdef NOTYET
3540 			/*
3541 			 * XXX jonathan@NetBSD.org: untested.
3542 			 * how to force  this branch to be taken?
3543 			 */
3544 			BGE_EVCNT_INCR(&sc->sc_ev_txtsopain);
3545 
3546 			m_copydata(m0, offset, sizeof(ip), &ip);
3547 			m_copydata(m0, hlen, sizeof(th), &th);
3548 
3549 			ip.ip_len = 0;
3550 
3551 			m_copyback(m0, hlen + offsetof(struct ip, ip_len),
3552 			    sizeof(ip.ip_len), &ip.ip_len);
3553 
3554 			th.th_sum = in_cksum_phdr(ip.ip_src.s_addr,
3555 			    ip.ip_dst.s_addr, htons(IPPROTO_TCP));
3556 
3557 			m_copyback(m0, hlen + offsetof(struct tcphdr, th_sum),
3558 			    sizeof(th.th_sum), &th.th_sum);
3559 
3560 			hlen += th.th_off << 2;
3561 			iptcp_opt_words	= hlen;
3562 #else
3563 			/*
3564 			 * if_wm "hard" case not yet supported, can we not
3565 			 * mandate it out of existence?
3566 			 */
3567 			(void) ip; (void)th; (void) ip_tcp_hlen;
3568 
3569 			return ENOBUFS;
3570 #endif
3571 		} else {
3572 			ip = (struct ip *) (mtod(m0, caddr_t) + offset);
3573 			th = (struct tcphdr *) (mtod(m0, caddr_t) + hlen);
3574 			ip_tcp_hlen = iphl +  (th->th_off << 2);
3575 
3576 			/* Total IP/TCP options, in 32-bit words */
3577 			iptcp_opt_words = (ip_tcp_hlen
3578 					   - sizeof(struct tcphdr)
3579 					   - sizeof(struct ip)) >> 2;
3580 		}
3581 		if (BGE_IS_5750_OR_BEYOND(sc)) {
3582 			th->th_sum = 0;
3583 			csum_flags &= ~(BGE_TXBDFLAG_TCP_UDP_CSUM);
3584 		} else {
3585 			/*
3586 			 * XXX jonathan@NetBSD.org: 5705 untested.
3587 			 * Requires TSO firmware patch for 5701/5703/5704.
3588 			 */
3589 			th->th_sum = in_cksum_phdr(ip->ip_src.s_addr,
3590 			    ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3591 		}
3592 
3593 		mss = m_head->m_pkthdr.segsz;
3594 		txbd_tso_flags |=
3595 		    BGE_TXBDFLAG_CPU_PRE_DMA |
3596 		    BGE_TXBDFLAG_CPU_POST_DMA;
3597 
3598 		/*
3599 		 * Our NIC TSO-assist assumes TSO has standard, optionless
3600 		 * IPv4 and TCP headers, which total 40 bytes. By default,
3601 		 * the NIC copies 40 bytes of IP/TCP header from the
3602 		 * supplied header into the IP/TCP header portion of
3603 		 * each post-TSO-segment. If the supplied packet has IP or
3604 		 * TCP options, we need to tell the NIC to copy those extra
3605 		 * bytes into each  post-TSO header, in addition to the normal
3606 		 * 40-byte IP/TCP header (and to leave space accordingly).
3607 		 * Unfortunately, the driver encoding of option length
3608 		 * varies across different ASIC families.
3609 		 */
3610 		tcp_seg_flags = 0;
3611 		if (iptcp_opt_words) {
3612 			if ( BGE_IS_5705_OR_BEYOND(sc)) {
3613 				tcp_seg_flags =
3614 					iptcp_opt_words << 11;
3615 			} else {
3616 				txbd_tso_flags |=
3617 					iptcp_opt_words << 12;
3618 			}
3619 		}
3620 		maxsegsize = mss | tcp_seg_flags;
3621 		ip->ip_len = htons(mss + ip_tcp_hlen);
3622 
3623 	}	/* TSO setup */
3624 
3625 	/*
3626 	 * Start packing the mbufs in this chain into
3627 	 * the fragment pointers. Stop when we run out
3628 	 * of fragments or hit the end of the mbuf chain.
3629 	 */
3630 	error = bus_dmamap_load_mbuf(sc->bge_dmatag, dmamap, m_head,
3631 	    BUS_DMA_NOWAIT);
3632 	if (error) {
3633 		return(ENOBUFS);
3634 	}
3635 
3636 	mtag = sc->ethercom.ec_nvlans ?
3637 	    m_tag_find(m_head, PACKET_TAG_VLAN, NULL) : NULL;
3638 
3639 
3640 	/* Iterate over dmap-map fragments. */
3641 	for (i = 0; i < dmamap->dm_nsegs; i++) {
3642 		f = &sc->bge_rdata->bge_tx_ring[frag];
3643 		if (sc->bge_cdata.bge_tx_chain[frag] != NULL)
3644 			break;
3645 
3646 		bge_set_hostaddr(&f->bge_addr, dmamap->dm_segs[i].ds_addr);
3647 		f->bge_len = dmamap->dm_segs[i].ds_len;
3648 
3649 		/*
3650 		 * For 5751 and follow-ons, for TSO we must turn
3651 		 * off checksum-assist flag in the tx-descr, and
3652 		 * supply the ASIC-revision-specific encoding
3653 		 * of TSO flags and segsize.
3654 		 */
3655 		if (use_tso) {
3656 			if (BGE_IS_5750_OR_BEYOND(sc) || i == 0) {
3657 				f->bge_rsvd = maxsegsize;
3658 				f->bge_flags = csum_flags | txbd_tso_flags;
3659 			} else {
3660 				f->bge_rsvd = 0;
3661 				f->bge_flags =
3662 				  (csum_flags | txbd_tso_flags) & 0x0fff;
3663 			}
3664 		} else {
3665 			f->bge_rsvd = 0;
3666 			f->bge_flags = csum_flags;
3667 		}
3668 
3669 		if (mtag != NULL) {
3670 			f->bge_flags |= BGE_TXBDFLAG_VLAN_TAG;
3671 			f->bge_vlan_tag = VLAN_TAG_VALUE(mtag);
3672 		} else {
3673 			f->bge_vlan_tag = 0;
3674 		}
3675 		/*
3676 		 * Sanity check: avoid coming within 16 descriptors
3677 		 * of the end of the ring.
3678 		 */
3679 		if ((BGE_TX_RING_CNT - (sc->bge_txcnt + cnt)) < 16) {
3680 			BGE_TSO_PRINTF(("%s: "
3681 			    " dmamap_load_mbuf too close to ring wrap\n",
3682 			    sc->bge_dev.dv_xname));
3683 			return(ENOBUFS);
3684 		}
3685 		cur = frag;
3686 		BGE_INC(frag, BGE_TX_RING_CNT);
3687 		cnt++;
3688 	}
3689 
3690 	if (i < dmamap->dm_nsegs) {
3691 		BGE_TSO_PRINTF(("%s: reached %d < dm_nsegs %d\n",
3692 		    sc->bge_dev.dv_xname, i, dmamap->dm_nsegs));
3693 		return ENOBUFS;
3694 	}
3695 
3696 	bus_dmamap_sync(sc->bge_dmatag, dmamap, 0, dmamap->dm_mapsize,
3697 	    BUS_DMASYNC_PREWRITE);
3698 
3699 	if (frag == sc->bge_tx_saved_considx) {
3700 		BGE_TSO_PRINTF(("%s: frag %d = wrapped id %d?\n",
3701 		    sc->bge_dev.dv_xname, frag, sc->bge_tx_saved_considx));
3702 
3703 		return(ENOBUFS);
3704 	}
3705 
3706 	sc->bge_rdata->bge_tx_ring[cur].bge_flags |= BGE_TXBDFLAG_END;
3707 	sc->bge_cdata.bge_tx_chain[cur] = m_head;
3708 	SLIST_REMOVE_HEAD(&sc->txdma_list, link);
3709 	sc->txdma[cur] = dma;
3710 	sc->bge_txcnt += cnt;
3711 
3712 	*txidx = frag;
3713 
3714 	return(0);
3715 }
3716 
3717 /*
3718  * Main transmit routine. To avoid having to do mbuf copies, we put pointers
3719  * to the mbuf data regions directly in the transmit descriptors.
3720  */
3721 static void
3722 bge_start(struct ifnet *ifp)
3723 {
3724 	struct bge_softc *sc;
3725 	struct mbuf *m_head = NULL;
3726 	u_int32_t prodidx;
3727 	int pkts = 0;
3728 
3729 	sc = ifp->if_softc;
3730 
3731 	if (!sc->bge_link && ifp->if_snd.ifq_len < 10)
3732 		return;
3733 
3734 	prodidx = sc->bge_tx_prodidx;
3735 
3736 	while(sc->bge_cdata.bge_tx_chain[prodidx] == NULL) {
3737 		IFQ_POLL(&ifp->if_snd, m_head);
3738 		if (m_head == NULL)
3739 			break;
3740 
3741 #if 0
3742 		/*
3743 		 * XXX
3744 		 * safety overkill.  If this is a fragmented packet chain
3745 		 * with delayed TCP/UDP checksums, then only encapsulate
3746 		 * it if we have enough descriptors to handle the entire
3747 		 * chain at once.
3748 		 * (paranoia -- may not actually be needed)
3749 		 */
3750 		if (m_head->m_flags & M_FIRSTFRAG &&
3751 		    m_head->m_pkthdr.csum_flags & (CSUM_DELAY_DATA)) {
3752 			if ((BGE_TX_RING_CNT - sc->bge_txcnt) <
3753 			    M_CSUM_DATA_IPv4_OFFSET(m_head->m_pkthdr.csum_data) + 16) {
3754 				ifp->if_flags |= IFF_OACTIVE;
3755 				break;
3756 			}
3757 		}
3758 #endif
3759 
3760 		/*
3761 		 * Pack the data into the transmit ring. If we
3762 		 * don't have room, set the OACTIVE flag and wait
3763 		 * for the NIC to drain the ring.
3764 		 */
3765 		if (bge_encap(sc, m_head, &prodidx)) {
3766 			printf("bge: failed on len %d?\n", m_head->m_pkthdr.len);
3767 			ifp->if_flags |= IFF_OACTIVE;
3768 			break;
3769 		}
3770 
3771 		/* now we are committed to transmit the packet */
3772 		IFQ_DEQUEUE(&ifp->if_snd, m_head);
3773 		pkts++;
3774 
3775 #if NBPFILTER > 0
3776 		/*
3777 		 * If there's a BPF listener, bounce a copy of this frame
3778 		 * to him.
3779 		 */
3780 		if (ifp->if_bpf)
3781 			bpf_mtap(ifp->if_bpf, m_head);
3782 #endif
3783 	}
3784 	if (pkts == 0)
3785 		return;
3786 
3787 	/* Transmit */
3788 	CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
3789 	if (sc->bge_quirks & BGE_QUIRK_PRODUCER_BUG)	/* 5700 b2 errata */
3790 		CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
3791 
3792 	sc->bge_tx_prodidx = prodidx;
3793 
3794 	/*
3795 	 * Set a timeout in case the chip goes out to lunch.
3796 	 */
3797 	ifp->if_timer = 5;
3798 }
3799 
3800 static int
3801 bge_init(struct ifnet *ifp)
3802 {
3803 	struct bge_softc *sc = ifp->if_softc;
3804 	u_int16_t *m;
3805 	int s, error;
3806 
3807 	s = splnet();
3808 
3809 	ifp = &sc->ethercom.ec_if;
3810 
3811 	/* Cancel pending I/O and flush buffers. */
3812 	bge_stop(sc);
3813 	bge_reset(sc);
3814 	bge_chipinit(sc);
3815 
3816 	/*
3817 	 * Init the various state machines, ring
3818 	 * control blocks and firmware.
3819 	 */
3820 	error = bge_blockinit(sc);
3821 	if (error != 0) {
3822 		printf("%s: initialization error %d\n", sc->bge_dev.dv_xname,
3823 		    error);
3824 		splx(s);
3825 		return error;
3826 	}
3827 
3828 	ifp = &sc->ethercom.ec_if;
3829 
3830 	/* Specify MTU. */
3831 	CSR_WRITE_4(sc, BGE_RX_MTU, ifp->if_mtu +
3832 	    ETHER_HDR_LEN + ETHER_CRC_LEN + ETHER_VLAN_ENCAP_LEN);
3833 
3834 	/* Load our MAC address. */
3835 	m = (u_int16_t *)&(LLADDR(ifp->if_sadl)[0]);
3836 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_LO, htons(m[0]));
3837 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_HI, (htons(m[1]) << 16) | htons(m[2]));
3838 
3839 	/* Enable or disable promiscuous mode as needed. */
3840 	if (ifp->if_flags & IFF_PROMISC) {
3841 		BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
3842 	} else {
3843 		BGE_CLRBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
3844 	}
3845 
3846 	/* Program multicast filter. */
3847 	bge_setmulti(sc);
3848 
3849 	/* Init RX ring. */
3850 	bge_init_rx_ring_std(sc);
3851 
3852 	/* Init jumbo RX ring. */
3853 	if (ifp->if_mtu > (ETHERMTU + ETHER_HDR_LEN + ETHER_CRC_LEN))
3854 		bge_init_rx_ring_jumbo(sc);
3855 
3856 	/* Init our RX return ring index */
3857 	sc->bge_rx_saved_considx = 0;
3858 
3859 	/* Init TX ring. */
3860 	bge_init_tx_ring(sc);
3861 
3862 	/* Turn on transmitter */
3863 	BGE_SETBIT(sc, BGE_TX_MODE, BGE_TXMODE_ENABLE);
3864 
3865 	/* Turn on receiver */
3866 	BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
3867 
3868 	CSR_WRITE_4(sc, BGE_MAX_RX_FRAME_LOWAT, 2);
3869 
3870 	/* Tell firmware we're alive. */
3871 	BGE_SETBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
3872 
3873 	/* Enable host interrupts. */
3874 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_CLEAR_INTA);
3875 	BGE_CLRBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
3876 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 0);
3877 
3878 	bge_ifmedia_upd(ifp);
3879 
3880 	ifp->if_flags |= IFF_RUNNING;
3881 	ifp->if_flags &= ~IFF_OACTIVE;
3882 
3883 	splx(s);
3884 
3885 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
3886 
3887 	return 0;
3888 }
3889 
3890 /*
3891  * Set media options.
3892  */
3893 static int
3894 bge_ifmedia_upd(struct ifnet *ifp)
3895 {
3896 	struct bge_softc *sc = ifp->if_softc;
3897 	struct mii_data *mii = &sc->bge_mii;
3898 	struct ifmedia *ifm = &sc->bge_ifmedia;
3899 
3900 	/* If this is a 1000baseX NIC, enable the TBI port. */
3901 	if (sc->bge_tbi) {
3902 		if (IFM_TYPE(ifm->ifm_media) != IFM_ETHER)
3903 			return(EINVAL);
3904 		switch(IFM_SUBTYPE(ifm->ifm_media)) {
3905 		case IFM_AUTO:
3906 			break;
3907 		case IFM_1000_SX:
3908 			if ((ifm->ifm_media & IFM_GMASK) == IFM_FDX) {
3909 				BGE_CLRBIT(sc, BGE_MAC_MODE,
3910 				    BGE_MACMODE_HALF_DUPLEX);
3911 			} else {
3912 				BGE_SETBIT(sc, BGE_MAC_MODE,
3913 				    BGE_MACMODE_HALF_DUPLEX);
3914 			}
3915 			break;
3916 		default:
3917 			return(EINVAL);
3918 		}
3919 		/* XXX 802.3x flow control for 1000BASE-SX */
3920 		return(0);
3921 	}
3922 
3923 	sc->bge_link = 0;
3924 	mii_mediachg(mii);
3925 
3926 	return(0);
3927 }
3928 
3929 /*
3930  * Report current media status.
3931  */
3932 static void
3933 bge_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
3934 {
3935 	struct bge_softc *sc = ifp->if_softc;
3936 	struct mii_data *mii = &sc->bge_mii;
3937 
3938 	if (sc->bge_tbi) {
3939 		ifmr->ifm_status = IFM_AVALID;
3940 		ifmr->ifm_active = IFM_ETHER;
3941 		if (CSR_READ_4(sc, BGE_MAC_STS) &
3942 		    BGE_MACSTAT_TBI_PCS_SYNCHED)
3943 			ifmr->ifm_status |= IFM_ACTIVE;
3944 		ifmr->ifm_active |= IFM_1000_SX;
3945 		if (CSR_READ_4(sc, BGE_MAC_MODE) & BGE_MACMODE_HALF_DUPLEX)
3946 			ifmr->ifm_active |= IFM_HDX;
3947 		else
3948 			ifmr->ifm_active |= IFM_FDX;
3949 		return;
3950 	}
3951 
3952 	mii_pollstat(mii);
3953 	ifmr->ifm_status = mii->mii_media_status;
3954 	ifmr->ifm_active = (mii->mii_media_active & ~IFM_ETH_FMASK) |
3955 	    sc->bge_flowflags;
3956 }
3957 
3958 static int
3959 bge_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
3960 {
3961 	struct bge_softc *sc = ifp->if_softc;
3962 	struct ifreq *ifr = (struct ifreq *) data;
3963 	int s, error = 0;
3964 	struct mii_data *mii;
3965 
3966 	s = splnet();
3967 
3968 	switch(command) {
3969 	case SIOCSIFFLAGS:
3970 		if (ifp->if_flags & IFF_UP) {
3971 			/*
3972 			 * If only the state of the PROMISC flag changed,
3973 			 * then just use the 'set promisc mode' command
3974 			 * instead of reinitializing the entire NIC. Doing
3975 			 * a full re-init means reloading the firmware and
3976 			 * waiting for it to start up, which may take a
3977 			 * second or two.
3978 			 */
3979 			if (ifp->if_flags & IFF_RUNNING &&
3980 			    ifp->if_flags & IFF_PROMISC &&
3981 			    !(sc->bge_if_flags & IFF_PROMISC)) {
3982 				BGE_SETBIT(sc, BGE_RX_MODE,
3983 				    BGE_RXMODE_RX_PROMISC);
3984 			} else if (ifp->if_flags & IFF_RUNNING &&
3985 			    !(ifp->if_flags & IFF_PROMISC) &&
3986 			    sc->bge_if_flags & IFF_PROMISC) {
3987 				BGE_CLRBIT(sc, BGE_RX_MODE,
3988 				    BGE_RXMODE_RX_PROMISC);
3989 			} else if (!(sc->bge_if_flags & IFF_UP))
3990 				bge_init(ifp);
3991 		} else {
3992 			if (ifp->if_flags & IFF_RUNNING) {
3993 				bge_stop(sc);
3994 			}
3995 		}
3996 		sc->bge_if_flags = ifp->if_flags;
3997 		error = 0;
3998 		break;
3999 	case SIOCSIFMEDIA:
4000 		/* XXX Flow control is not supported for 1000BASE-SX */
4001 		if (sc->bge_tbi) {
4002 			ifr->ifr_media &= ~IFM_ETH_FMASK;
4003 			sc->bge_flowflags = 0;
4004 		}
4005 
4006 		/* Flow control requires full-duplex mode. */
4007 		if (IFM_SUBTYPE(ifr->ifr_media) == IFM_AUTO ||
4008 		    (ifr->ifr_media & IFM_FDX) == 0) {
4009 		    	ifr->ifr_media &= ~IFM_ETH_FMASK;
4010 		}
4011 		if (IFM_SUBTYPE(ifr->ifr_media) != IFM_AUTO) {
4012 			if ((ifr->ifr_media & IFM_ETH_FMASK) == IFM_FLOW) {
4013 				/* We an do both TXPAUSE and RXPAUSE. */
4014 				ifr->ifr_media |=
4015 				    IFM_ETH_TXPAUSE | IFM_ETH_RXPAUSE;
4016 			}
4017 			sc->bge_flowflags = ifr->ifr_media & IFM_ETH_FMASK;
4018 		}
4019 		/* FALLTHROUGH */
4020 	case SIOCGIFMEDIA:
4021 		if (sc->bge_tbi) {
4022 			error = ifmedia_ioctl(ifp, ifr, &sc->bge_ifmedia,
4023 			    command);
4024 		} else {
4025 			mii = &sc->bge_mii;
4026 			error = ifmedia_ioctl(ifp, ifr, &mii->mii_media,
4027 			    command);
4028 		}
4029 		break;
4030 	default:
4031 		error = ether_ioctl(ifp, command, data);
4032 		if (error == ENETRESET) {
4033 			if (ifp->if_flags & IFF_RUNNING)
4034 				bge_setmulti(sc);
4035 			error = 0;
4036 		}
4037 		break;
4038 	}
4039 
4040 	splx(s);
4041 
4042 	return(error);
4043 }
4044 
4045 static void
4046 bge_watchdog(struct ifnet *ifp)
4047 {
4048 	struct bge_softc *sc;
4049 
4050 	sc = ifp->if_softc;
4051 
4052 	printf("%s: watchdog timeout -- resetting\n", sc->bge_dev.dv_xname);
4053 
4054 	ifp->if_flags &= ~IFF_RUNNING;
4055 	bge_init(ifp);
4056 
4057 	ifp->if_oerrors++;
4058 }
4059 
4060 static void
4061 bge_stop_block(struct bge_softc *sc, bus_addr_t reg, uint32_t bit)
4062 {
4063 	int i;
4064 
4065 	BGE_CLRBIT(sc, reg, bit);
4066 
4067 	for (i = 0; i < BGE_TIMEOUT; i++) {
4068 		if ((CSR_READ_4(sc, reg) & bit) == 0)
4069 			return;
4070 		delay(100);
4071 		if (sc->bge_pcie)
4072 		  DELAY(1000);
4073 	}
4074 
4075 	printf("%s: block failed to stop: reg 0x%lx, bit 0x%08x\n",
4076 	    sc->bge_dev.dv_xname, (u_long) reg, bit);
4077 }
4078 
4079 /*
4080  * Stop the adapter and free any mbufs allocated to the
4081  * RX and TX lists.
4082  */
4083 static void
4084 bge_stop(struct bge_softc *sc)
4085 {
4086 	struct ifnet *ifp = &sc->ethercom.ec_if;
4087 
4088 	callout_stop(&sc->bge_timeout);
4089 
4090 	/*
4091 	 * Disable all of the receiver blocks
4092 	 */
4093 	bge_stop_block(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
4094 	bge_stop_block(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
4095 	bge_stop_block(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
4096 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4097 		bge_stop_block(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
4098 	}
4099 	bge_stop_block(sc, BGE_RDBDI_MODE, BGE_RBDIMODE_ENABLE);
4100 	bge_stop_block(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
4101 	bge_stop_block(sc, BGE_RBDC_MODE, BGE_RBDCMODE_ENABLE);
4102 
4103 	/*
4104 	 * Disable all of the transmit blocks
4105 	 */
4106 	bge_stop_block(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
4107 	bge_stop_block(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
4108 	bge_stop_block(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
4109 	bge_stop_block(sc, BGE_RDMA_MODE, BGE_RDMAMODE_ENABLE);
4110 	bge_stop_block(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
4111 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4112 		bge_stop_block(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
4113 	}
4114 	bge_stop_block(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
4115 
4116 	/*
4117 	 * Shut down all of the memory managers and related
4118 	 * state machines.
4119 	 */
4120 	bge_stop_block(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
4121 	bge_stop_block(sc, BGE_WDMA_MODE, BGE_WDMAMODE_ENABLE);
4122 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4123 		bge_stop_block(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
4124 	}
4125 
4126 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0xFFFFFFFF);
4127 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0);
4128 
4129 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4130 		bge_stop_block(sc, BGE_BMAN_MODE, BGE_BMANMODE_ENABLE);
4131 		bge_stop_block(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE);
4132 	}
4133 
4134 	/* Disable host interrupts. */
4135 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
4136 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 1);
4137 
4138 	/*
4139 	 * Tell firmware we're shutting down.
4140 	 */
4141 	BGE_CLRBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
4142 
4143 	/* Free the RX lists. */
4144 	bge_free_rx_ring_std(sc);
4145 
4146 	/* Free jumbo RX list. */
4147 	bge_free_rx_ring_jumbo(sc);
4148 
4149 	/* Free TX buffers. */
4150 	bge_free_tx_ring(sc);
4151 
4152 	/*
4153 	 * Isolate/power down the PHY.
4154 	 */
4155 	if (!sc->bge_tbi)
4156 		mii_down(&sc->bge_mii);
4157 
4158 	sc->bge_link = 0;
4159 
4160 	sc->bge_tx_saved_considx = BGE_TXCONS_UNSET;
4161 
4162 	ifp->if_flags &= ~(IFF_RUNNING | IFF_OACTIVE);
4163 }
4164 
4165 /*
4166  * Stop all chip I/O so that the kernel's probe routines don't
4167  * get confused by errant DMAs when rebooting.
4168  */
4169 static void
4170 bge_shutdown(void *xsc)
4171 {
4172 	struct bge_softc *sc = (struct bge_softc *)xsc;
4173 
4174 	bge_stop(sc);
4175 	bge_reset(sc);
4176 }
4177 
4178 
4179 static int
4180 sysctl_bge_verify(SYSCTLFN_ARGS)
4181 {
4182 	int error, t;
4183 	struct sysctlnode node;
4184 
4185 	node = *rnode;
4186 	t = *(int*)rnode->sysctl_data;
4187 	node.sysctl_data = &t;
4188 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
4189 	if (error || newp == NULL)
4190 		return (error);
4191 
4192 #if 0
4193 	DPRINTF2(("%s: t = %d, nodenum = %d, rnodenum = %d\n", __func__, t,
4194 	    node.sysctl_num, rnode->sysctl_num));
4195 #endif
4196 
4197 	if (node.sysctl_num == bge_rxthresh_nodenum) {
4198 		if (t < 0 || t >= NBGE_RX_THRESH)
4199 			return (EINVAL);
4200 		bge_update_all_threshes(t);
4201 	} else
4202 		return (EINVAL);
4203 
4204 	*(int*)rnode->sysctl_data = t;
4205 
4206 	return (0);
4207 }
4208 
4209 /*
4210  * Set up sysctl(3) MIB, hw.bge.*.
4211  *
4212  * TBD condition SYSCTL_PERMANENT on being an LKM or not
4213  */
4214 SYSCTL_SETUP(sysctl_bge, "sysctl bge subtree setup")
4215 {
4216 	int rc, bge_root_num;
4217 	const struct sysctlnode *node;
4218 
4219 	if ((rc = sysctl_createv(clog, 0, NULL, NULL,
4220 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "hw", NULL,
4221 	    NULL, 0, NULL, 0, CTL_HW, CTL_EOL)) != 0) {
4222 		goto err;
4223 	}
4224 
4225 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
4226 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "bge",
4227 	    SYSCTL_DESCR("BGE interface controls"),
4228 	    NULL, 0, NULL, 0, CTL_HW, CTL_CREATE, CTL_EOL)) != 0) {
4229 		goto err;
4230 	}
4231 
4232 	bge_root_num = node->sysctl_num;
4233 
4234 	/* BGE Rx interrupt mitigation level */
4235 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
4236 	    CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
4237 	    CTLTYPE_INT, "rx_lvl",
4238 	    SYSCTL_DESCR("BGE receive interrupt mitigation level"),
4239 	    sysctl_bge_verify, 0,
4240 	    &bge_rx_thresh_lvl,
4241 	    0, CTL_HW, bge_root_num, CTL_CREATE,
4242 	    CTL_EOL)) != 0) {
4243 		goto err;
4244 	}
4245 
4246 	bge_rxthresh_nodenum = node->sysctl_num;
4247 
4248 	return;
4249 
4250 err:
4251 	printf("%s: sysctl_createv failed (rc = %d)\n", __func__, rc);
4252 }
4253 
4254 static void
4255 bge_powerhook(int why, void *hdl)
4256 {
4257 	struct bge_softc *sc = (struct bge_softc *)hdl;
4258 	struct ifnet *ifp = &sc->ethercom.ec_if;
4259 	struct pci_attach_args *pa = &(sc->bge_pa);
4260 	pci_chipset_tag_t pc = pa->pa_pc;
4261 	pcitag_t tag = pa->pa_tag;
4262 
4263 	switch (why) {
4264 	case PWR_SOFTSUSPEND:
4265 	case PWR_SOFTSTANDBY:
4266 		bge_shutdown(sc);
4267 		break;
4268 	case PWR_SOFTRESUME:
4269 		if (ifp->if_flags & IFF_UP) {
4270 			ifp->if_flags &= ~IFF_RUNNING;
4271 			bge_init(ifp);
4272 		}
4273 		break;
4274 	case PWR_SUSPEND:
4275 	case PWR_STANDBY:
4276 		pci_conf_capture(pc, tag, &sc->bge_pciconf);
4277 		break;
4278 	case PWR_RESUME:
4279 		pci_conf_restore(pc, tag, &sc->bge_pciconf);
4280 		break;
4281 	}
4282 
4283 	return;
4284 }
4285