xref: /netbsd-src/sys/dev/pci/if_bge.c (revision c71562d660be5e4ad22016bce45e96f08af190cc)
1 /*	$NetBSD: if_bge.c,v 1.105 2006/04/18 14:53:43 christos 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 1.105 2006/04/18 14:53:43 christos 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 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1488 		CSR_WRITE_4(sc, BGE_BMAN_MODE,
1489 		    BGE_BMANMODE_ENABLE|BGE_BMANMODE_LOMBUF_ATTN);
1490 
1491 		/* Poll for buffer manager start indication */
1492 		for (i = 0; i < BGE_TIMEOUT; i++) {
1493 			if (CSR_READ_4(sc, BGE_BMAN_MODE) & BGE_BMANMODE_ENABLE)
1494 				break;
1495 			DELAY(10);
1496 		}
1497 
1498 		if (i == BGE_TIMEOUT) {
1499 			printf("%s: buffer manager failed to start\n",
1500 			    sc->bge_dev.dv_xname);
1501 			return(ENXIO);
1502 		}
1503 	}
1504 
1505 	/* Enable flow-through queues */
1506 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0xFFFFFFFF);
1507 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0);
1508 
1509 	/* Wait until queue initialization is complete */
1510 	for (i = 0; i < BGE_TIMEOUT; i++) {
1511 		if (CSR_READ_4(sc, BGE_FTQ_RESET) == 0)
1512 			break;
1513 		DELAY(10);
1514 	}
1515 
1516 	if (i == BGE_TIMEOUT) {
1517 		printf("%s: flow-through queue init failed\n",
1518 		    sc->bge_dev.dv_xname);
1519 		return(ENXIO);
1520 	}
1521 
1522 	/* Initialize the standard RX ring control block */
1523 	rcb = &sc->bge_rdata->bge_info.bge_std_rx_rcb;
1524 	bge_set_hostaddr(&rcb->bge_hostaddr,
1525 	    BGE_RING_DMA_ADDR(sc, bge_rx_std_ring));
1526 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1527 		rcb->bge_maxlen_flags =
1528 		    BGE_RCB_MAXLEN_FLAGS(BGE_MAX_FRAMELEN, 0);
1529 	} else {
1530 		rcb->bge_maxlen_flags = BGE_RCB_MAXLEN_FLAGS(512, 0);
1531 	}
1532 	if (sc->bge_extram)
1533 		rcb->bge_nicaddr = BGE_EXT_STD_RX_RINGS;
1534 	else
1535 		rcb->bge_nicaddr = BGE_STD_RX_RINGS;
1536 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_HADDR_HI, rcb->bge_hostaddr.bge_addr_hi);
1537 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_HADDR_LO, rcb->bge_hostaddr.bge_addr_lo);
1538 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_MAXLEN_FLAGS, rcb->bge_maxlen_flags);
1539 	CSR_WRITE_4(sc, BGE_RX_STD_RCB_NICADDR, rcb->bge_nicaddr);
1540 
1541 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1542 		sc->bge_return_ring_cnt = BGE_RETURN_RING_CNT;
1543 	} else {
1544 		sc->bge_return_ring_cnt = BGE_RETURN_RING_CNT_5705;
1545 	}
1546 
1547 	/*
1548 	 * Initialize the jumbo RX ring control block
1549 	 * We set the 'ring disabled' bit in the flags
1550 	 * field until we're actually ready to start
1551 	 * using this ring (i.e. once we set the MTU
1552 	 * high enough to require it).
1553 	 */
1554 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1555 		rcb = &sc->bge_rdata->bge_info.bge_jumbo_rx_rcb;
1556 		bge_set_hostaddr(&rcb->bge_hostaddr,
1557 		    BGE_RING_DMA_ADDR(sc, bge_rx_jumbo_ring));
1558 		rcb->bge_maxlen_flags =
1559 		    BGE_RCB_MAXLEN_FLAGS(BGE_MAX_FRAMELEN,
1560 			BGE_RCB_FLAG_RING_DISABLED);
1561 		if (sc->bge_extram)
1562 			rcb->bge_nicaddr = BGE_EXT_JUMBO_RX_RINGS;
1563 		else
1564 			rcb->bge_nicaddr = BGE_JUMBO_RX_RINGS;
1565 
1566 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_HADDR_HI,
1567 		    rcb->bge_hostaddr.bge_addr_hi);
1568 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_HADDR_LO,
1569 		    rcb->bge_hostaddr.bge_addr_lo);
1570 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_MAXLEN_FLAGS,
1571 		    rcb->bge_maxlen_flags);
1572 		CSR_WRITE_4(sc, BGE_RX_JUMBO_RCB_NICADDR, rcb->bge_nicaddr);
1573 
1574 		/* Set up dummy disabled mini ring RCB */
1575 		rcb = &sc->bge_rdata->bge_info.bge_mini_rx_rcb;
1576 		rcb->bge_maxlen_flags = BGE_RCB_MAXLEN_FLAGS(0,
1577 		    BGE_RCB_FLAG_RING_DISABLED);
1578 		CSR_WRITE_4(sc, BGE_RX_MINI_RCB_MAXLEN_FLAGS,
1579 		    rcb->bge_maxlen_flags);
1580 
1581 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
1582 		    offsetof(struct bge_ring_data, bge_info),
1583 		    sizeof (struct bge_gib),
1584 		    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1585 	}
1586 
1587 	/*
1588 	 * Set the BD ring replentish thresholds. The recommended
1589 	 * values are 1/8th the number of descriptors allocated to
1590 	 * each ring.
1591 	 */
1592 	CSR_WRITE_4(sc, BGE_RBDI_STD_REPL_THRESH, BGE_STD_RX_RING_CNT/8);
1593 	CSR_WRITE_4(sc, BGE_RBDI_JUMBO_REPL_THRESH, BGE_JUMBO_RX_RING_CNT/8);
1594 
1595 	/*
1596 	 * Disable all unused send rings by setting the 'ring disabled'
1597 	 * bit in the flags field of all the TX send ring control blocks.
1598 	 * These are located in NIC memory.
1599 	 */
1600 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
1601 	for (i = 0; i < BGE_TX_RINGS_EXTSSRAM_MAX; i++) {
1602 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1603 		    BGE_RCB_MAXLEN_FLAGS(0,BGE_RCB_FLAG_RING_DISABLED));
1604 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
1605 		rcb_addr += sizeof(struct bge_rcb);
1606 	}
1607 
1608 	/* Configure TX RCB 0 (we use only the first ring) */
1609 	rcb_addr = BGE_MEMWIN_START + BGE_SEND_RING_RCB;
1610 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_tx_ring));
1611 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
1612 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
1613 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr,
1614 		    BGE_NIC_TXRING_ADDR(0, BGE_TX_RING_CNT));
1615 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1616 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1617 		    BGE_RCB_MAXLEN_FLAGS(BGE_TX_RING_CNT, 0));
1618 	}
1619 
1620 	/* Disable all unused RX return rings */
1621 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
1622 	for (i = 0; i < BGE_RX_RINGS_MAX; i++) {
1623 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, 0);
1624 		RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, 0);
1625 		RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1626 			    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt,
1627                                      BGE_RCB_FLAG_RING_DISABLED));
1628 		RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0);
1629 		CSR_WRITE_4(sc, BGE_MBX_RX_CONS0_LO +
1630 		    (i * (sizeof(u_int64_t))), 0);
1631 		rcb_addr += sizeof(struct bge_rcb);
1632 	}
1633 
1634 	/* Initialize RX ring indexes */
1635 	CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, 0);
1636 	CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, 0);
1637 	CSR_WRITE_4(sc, BGE_MBX_RX_MINI_PROD_LO, 0);
1638 
1639 	/*
1640 	 * Set up RX return ring 0
1641 	 * Note that the NIC address for RX return rings is 0x00000000.
1642 	 * The return rings live entirely within the host, so the
1643 	 * nicaddr field in the RCB isn't used.
1644 	 */
1645 	rcb_addr = BGE_MEMWIN_START + BGE_RX_RETURN_RING_RCB;
1646 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_rx_return_ring));
1647 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_hi, taddr.bge_addr_hi);
1648 	RCB_WRITE_4(sc, rcb_addr, bge_hostaddr.bge_addr_lo, taddr.bge_addr_lo);
1649 	RCB_WRITE_4(sc, rcb_addr, bge_nicaddr, 0x00000000);
1650 	RCB_WRITE_4(sc, rcb_addr, bge_maxlen_flags,
1651 	    BGE_RCB_MAXLEN_FLAGS(sc->bge_return_ring_cnt, 0));
1652 
1653 	/* Set random backoff seed for TX */
1654 	CSR_WRITE_4(sc, BGE_TX_RANDOM_BACKOFF,
1655 	    LLADDR(ifp->if_sadl)[0] + LLADDR(ifp->if_sadl)[1] +
1656 	    LLADDR(ifp->if_sadl)[2] + LLADDR(ifp->if_sadl)[3] +
1657 	    LLADDR(ifp->if_sadl)[4] + LLADDR(ifp->if_sadl)[5] +
1658 	    BGE_TX_BACKOFF_SEED_MASK);
1659 
1660 	/* Set inter-packet gap */
1661 	CSR_WRITE_4(sc, BGE_TX_LENGTHS, 0x2620);
1662 
1663 	/*
1664 	 * Specify which ring to use for packets that don't match
1665 	 * any RX rules.
1666 	 */
1667 	CSR_WRITE_4(sc, BGE_RX_RULES_CFG, 0x08);
1668 
1669 	/*
1670 	 * Configure number of RX lists. One interrupt distribution
1671 	 * list, sixteen active lists, one bad frames class.
1672 	 */
1673 	CSR_WRITE_4(sc, BGE_RXLP_CFG, 0x181);
1674 
1675 	/* Inialize RX list placement stats mask. */
1676 	CSR_WRITE_4(sc, BGE_RXLP_STATS_ENABLE_MASK, 0x007FFFFF);
1677 	CSR_WRITE_4(sc, BGE_RXLP_STATS_CTL, 0x1);
1678 
1679 	/* Disable host coalescing until we get it set up */
1680 	CSR_WRITE_4(sc, BGE_HCC_MODE, 0x00000000);
1681 
1682 	/* Poll to make sure it's shut down. */
1683 	for (i = 0; i < BGE_TIMEOUT; i++) {
1684 		if (!(CSR_READ_4(sc, BGE_HCC_MODE) & BGE_HCCMODE_ENABLE))
1685 			break;
1686 		DELAY(10);
1687 	}
1688 
1689 	if (i == BGE_TIMEOUT) {
1690 		printf("%s: host coalescing engine failed to idle\n",
1691 		    sc->bge_dev.dv_xname);
1692 		return(ENXIO);
1693 	}
1694 
1695 	/* Set up host coalescing defaults */
1696 	CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, sc->bge_rx_coal_ticks);
1697 	CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS, sc->bge_tx_coal_ticks);
1698 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, sc->bge_rx_max_coal_bds);
1699 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS, sc->bge_tx_max_coal_bds);
1700 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1701 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS_INT, 0);
1702 		CSR_WRITE_4(sc, BGE_HCC_TX_COAL_TICKS_INT, 0);
1703 	}
1704 	CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS_INT, 0);
1705 	CSR_WRITE_4(sc, BGE_HCC_TX_MAX_COAL_BDS_INT, 0);
1706 
1707 	/* Set up address of statistics block */
1708 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1709 		bge_set_hostaddr(&taddr,
1710 		    BGE_RING_DMA_ADDR(sc, bge_info.bge_stats));
1711 		CSR_WRITE_4(sc, BGE_HCC_STATS_TICKS, sc->bge_stat_ticks);
1712 		CSR_WRITE_4(sc, BGE_HCC_STATS_BASEADDR, BGE_STATS_BLOCK);
1713 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_HI, taddr.bge_addr_hi);
1714 		CSR_WRITE_4(sc, BGE_HCC_STATS_ADDR_LO, taddr.bge_addr_lo);
1715 	}
1716 
1717 	/* Set up address of status block */
1718 	bge_set_hostaddr(&taddr, BGE_RING_DMA_ADDR(sc, bge_status_block));
1719 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_BASEADDR, BGE_STATUS_BLOCK);
1720 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_HI, taddr.bge_addr_hi);
1721 	CSR_WRITE_4(sc, BGE_HCC_STATUSBLK_ADDR_LO, taddr.bge_addr_lo);
1722 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx = 0;
1723 	sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx = 0;
1724 
1725 	/* Turn on host coalescing state machine */
1726 	CSR_WRITE_4(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
1727 
1728 	/* Turn on RX BD completion state machine and enable attentions */
1729 	CSR_WRITE_4(sc, BGE_RBDC_MODE,
1730 	    BGE_RBDCMODE_ENABLE|BGE_RBDCMODE_ATTN);
1731 
1732 	/* Turn on RX list placement state machine */
1733 	CSR_WRITE_4(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
1734 
1735 	/* Turn on RX list selector state machine. */
1736 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1737 		CSR_WRITE_4(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
1738 	}
1739 
1740 	/* Turn on DMA, clear stats */
1741 	CSR_WRITE_4(sc, BGE_MAC_MODE, BGE_MACMODE_TXDMA_ENB|
1742 	    BGE_MACMODE_RXDMA_ENB|BGE_MACMODE_RX_STATS_CLEAR|
1743 	    BGE_MACMODE_TX_STATS_CLEAR|BGE_MACMODE_RX_STATS_ENB|
1744 	    BGE_MACMODE_TX_STATS_ENB|BGE_MACMODE_FRMHDR_DMA_ENB|
1745 	    (sc->bge_tbi ? BGE_PORTMODE_TBI : BGE_PORTMODE_MII));
1746 
1747 	/* Set misc. local control, enable interrupts on attentions */
1748 	sc->bge_local_ctrl_reg = BGE_MLC_INTR_ONATTN | BGE_MLC_AUTO_EEPROM;
1749 
1750 #ifdef notdef
1751 	/* Assert GPIO pins for PHY reset */
1752 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUT0|
1753 	    BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUT2);
1754 	BGE_SETBIT(sc, BGE_MISC_LOCAL_CTL, BGE_MLC_MISCIO_OUTEN0|
1755 	    BGE_MLC_MISCIO_OUTEN1|BGE_MLC_MISCIO_OUTEN2);
1756 #endif
1757 
1758 #if defined(not_quite_yet)
1759 	/* Linux driver enables enable gpio pin #1 on 5700s */
1760 	if (sc->bge_chipid == BGE_CHIPID_BCM5700) {
1761 		sc->bge_local_ctrl_reg |=
1762 		  (BGE_MLC_MISCIO_OUT1|BGE_MLC_MISCIO_OUTEN1);
1763 	}
1764 #endif
1765 	CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
1766 
1767 	/* Turn on DMA completion state machine */
1768 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1769 		CSR_WRITE_4(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
1770 	}
1771 
1772 	/* Turn on write DMA state machine */
1773 	CSR_WRITE_4(sc, BGE_WDMA_MODE,
1774 	    BGE_WDMAMODE_ENABLE|BGE_WDMAMODE_ALL_ATTNS);
1775 
1776 	/* Turn on read DMA state machine */
1777 	{
1778 		uint32_t dma_read_modebits;
1779 
1780 		dma_read_modebits =
1781 		  BGE_RDMAMODE_ENABLE | BGE_RDMAMODE_ALL_ATTNS;
1782 
1783 		if (sc->bge_pcie && 0) {
1784 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_LONG_BURST;
1785 		} else if ((sc->bge_quirks & BGE_QUIRK_5705_CORE)) {
1786 			dma_read_modebits |= BGE_RDMA_MODE_FIFO_SIZE_128;
1787 		}
1788 
1789 		/* XXX broadcom-supplied linux driver; undocumented */
1790 		if (BGE_IS_5750_OR_BEYOND(sc)) {
1791  			/*
1792 			 * XXX: magic values.
1793 			 * From Broadcom-supplied Linux driver;  apparently
1794 			 * required to workaround a DMA bug affecting TSO
1795 			 * on bcm575x/bcm5721?
1796 			 */
1797 			dma_read_modebits |= (1 << 27);
1798 		}
1799 		CSR_WRITE_4(sc, BGE_RDMA_MODE, dma_read_modebits);
1800 	}
1801 
1802 	/* Turn on RX data completion state machine */
1803 	CSR_WRITE_4(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
1804 
1805 	/* Turn on RX BD initiator state machine */
1806 	CSR_WRITE_4(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
1807 
1808 	/* Turn on RX data and RX BD initiator state machine */
1809 	CSR_WRITE_4(sc, BGE_RDBDI_MODE, BGE_RDBDIMODE_ENABLE);
1810 
1811 	/* Turn on Mbuf cluster free state machine */
1812 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
1813 		CSR_WRITE_4(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
1814 	}
1815 
1816 	/* Turn on send BD completion state machine */
1817 	CSR_WRITE_4(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
1818 
1819 	/* Turn on send data completion state machine */
1820 	CSR_WRITE_4(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
1821 
1822 	/* Turn on send data initiator state machine */
1823 	if (BGE_IS_5750_OR_BEYOND(sc)) {
1824 		/* XXX: magic value from Linux driver */
1825 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE | 0x08);
1826 	} else {
1827 		CSR_WRITE_4(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
1828 	}
1829 
1830 	/* Turn on send BD initiator state machine */
1831 	CSR_WRITE_4(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
1832 
1833 	/* Turn on send BD selector state machine */
1834 	CSR_WRITE_4(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
1835 
1836 	CSR_WRITE_4(sc, BGE_SDI_STATS_ENABLE_MASK, 0x007FFFFF);
1837 	CSR_WRITE_4(sc, BGE_SDI_STATS_CTL,
1838 	    BGE_SDISTATSCTL_ENABLE|BGE_SDISTATSCTL_FASTER);
1839 
1840 	/* ack/clear link change events */
1841 	CSR_WRITE_4(sc, BGE_MAC_STS, BGE_MACSTAT_SYNC_CHANGED|
1842 	    BGE_MACSTAT_CFG_CHANGED);
1843 	CSR_WRITE_4(sc, BGE_MI_STS, 0);
1844 
1845 	/* Enable PHY auto polling (for MII/GMII only) */
1846 	if (sc->bge_tbi) {
1847 		CSR_WRITE_4(sc, BGE_MI_STS, BGE_MISTS_LINK);
1848  	} else {
1849 		BGE_SETBIT(sc, BGE_MI_MODE, BGE_MIMODE_AUTOPOLL|10<<16);
1850 		if (sc->bge_quirks & BGE_QUIRK_LINK_STATE_BROKEN)
1851 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
1852 			    BGE_EVTENB_MI_INTERRUPT);
1853 	}
1854 
1855 	/* Enable link state change attentions. */
1856 	BGE_SETBIT(sc, BGE_MAC_EVT_ENB, BGE_EVTENB_LINK_CHANGED);
1857 
1858 	return(0);
1859 }
1860 
1861 static const struct bge_revision {
1862 	uint32_t		br_chipid;
1863 	uint32_t		br_quirks;
1864 	const char		*br_name;
1865 } bge_revisions[] = {
1866 	{ BGE_CHIPID_BCM5700_A0,
1867 	  BGE_QUIRK_LINK_STATE_BROKEN,
1868 	  "BCM5700 A0" },
1869 
1870 	{ BGE_CHIPID_BCM5700_A1,
1871 	  BGE_QUIRK_LINK_STATE_BROKEN,
1872 	  "BCM5700 A1" },
1873 
1874 	{ BGE_CHIPID_BCM5700_B0,
1875 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_CSUM_BROKEN|BGE_QUIRK_5700_COMMON,
1876 	  "BCM5700 B0" },
1877 
1878 	{ BGE_CHIPID_BCM5700_B1,
1879 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
1880 	  "BCM5700 B1" },
1881 
1882 	{ BGE_CHIPID_BCM5700_B2,
1883 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
1884 	  "BCM5700 B2" },
1885 
1886 	/* This is treated like a BCM5700 Bx */
1887 	{ BGE_CHIPID_BCM5700_ALTIMA,
1888 	  BGE_QUIRK_LINK_STATE_BROKEN|BGE_QUIRK_5700_COMMON,
1889 	  "BCM5700 Altima" },
1890 
1891 	{ BGE_CHIPID_BCM5700_C0,
1892 	  0,
1893 	  "BCM5700 C0" },
1894 
1895 	{ BGE_CHIPID_BCM5701_A0,
1896 	  0, /*XXX really, just not known */
1897 	  "BCM5701 A0" },
1898 
1899 	{ BGE_CHIPID_BCM5701_B0,
1900 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
1901 	  "BCM5701 B0" },
1902 
1903 	{ BGE_CHIPID_BCM5701_B2,
1904 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
1905 	  "BCM5701 B2" },
1906 
1907 	{ BGE_CHIPID_BCM5701_B5,
1908 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
1909 	  "BCM5701 B5" },
1910 
1911 	{ BGE_CHIPID_BCM5703_A0,
1912 	  0,
1913 	  "BCM5703 A0" },
1914 
1915 	{ BGE_CHIPID_BCM5703_A1,
1916 	  0,
1917 	  "BCM5703 A1" },
1918 
1919 	{ BGE_CHIPID_BCM5703_A2,
1920 	  BGE_QUIRK_ONLY_PHY_1,
1921 	  "BCM5703 A2" },
1922 
1923 	{ BGE_CHIPID_BCM5703_A3,
1924 	  BGE_QUIRK_ONLY_PHY_1,
1925 	  "BCM5703 A3" },
1926 
1927 	{ BGE_CHIPID_BCM5704_A0,
1928   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1929 	  "BCM5704 A0" },
1930 
1931 	{ BGE_CHIPID_BCM5704_A1,
1932   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1933 	  "BCM5704 A1" },
1934 
1935 	{ BGE_CHIPID_BCM5704_A2,
1936   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1937 	  "BCM5704 A2" },
1938 
1939 	{ BGE_CHIPID_BCM5704_A3,
1940   	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_FEWER_MBUFS,
1941 	  "BCM5704 A3" },
1942 
1943 	{ BGE_CHIPID_BCM5705_A0,
1944 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1945 	  "BCM5705 A0" },
1946 
1947 	{ BGE_CHIPID_BCM5705_A1,
1948 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1949 	  "BCM5705 A1" },
1950 
1951 	{ BGE_CHIPID_BCM5705_A2,
1952 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1953 	  "BCM5705 A2" },
1954 
1955 	{ BGE_CHIPID_BCM5705_A3,
1956 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1957 	  "BCM5705 A3" },
1958 
1959 	{ BGE_CHIPID_BCM5750_A0,
1960 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1961 	  "BCM5750 A1" },
1962 
1963 	{ BGE_CHIPID_BCM5750_A1,
1964 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1965 	  "BCM5750 A1" },
1966 
1967 	{ BGE_CHIPID_BCM5751_A1,
1968 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1969 	  "BCM5751 A1" },
1970 
1971 	{ 0, 0, NULL }
1972 };
1973 
1974 /*
1975  * Some defaults for major revisions, so that newer steppings
1976  * that we don't know about have a shot at working.
1977  */
1978 static const struct bge_revision bge_majorrevs[] = {
1979 	{ BGE_ASICREV_BCM5700,
1980 	  BGE_QUIRK_LINK_STATE_BROKEN,
1981 	  "unknown BCM5700" },
1982 
1983 	{ BGE_ASICREV_BCM5701,
1984 	  BGE_QUIRK_PCIX_DMA_ALIGN_BUG,
1985 	  "unknown BCM5701" },
1986 
1987 	{ BGE_ASICREV_BCM5703,
1988 	  0,
1989 	  "unknown BCM5703" },
1990 
1991 	{ BGE_ASICREV_BCM5704,
1992 	  BGE_QUIRK_ONLY_PHY_1,
1993 	  "unknown BCM5704" },
1994 
1995 	{ BGE_ASICREV_BCM5705,
1996 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
1997 	  "unknown BCM5705" },
1998 
1999 	{ BGE_ASICREV_BCM5750,
2000 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2001 	  "unknown BCM575x family" },
2002 
2003 	{ BGE_ASICREV_BCM5714,
2004 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2005 	  "unknown BCM5714" },
2006 
2007 	{ BGE_ASICREV_BCM5752,
2008 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2009 	  "unknown BCM5752 family" },
2010 
2011 
2012 	{ BGE_ASICREV_BCM5715,
2013 	  BGE_QUIRK_ONLY_PHY_1|BGE_QUIRK_5705_CORE,
2014 	  "unknown BCM5715" },
2015 
2016 	{ 0,
2017 	  0,
2018 	  NULL }
2019 };
2020 
2021 
2022 static const struct bge_revision *
2023 bge_lookup_rev(uint32_t chipid)
2024 {
2025 	const struct bge_revision *br;
2026 
2027 	for (br = bge_revisions; br->br_name != NULL; br++) {
2028 		if (br->br_chipid == chipid)
2029 			return (br);
2030 	}
2031 
2032 	for (br = bge_majorrevs; br->br_name != NULL; br++) {
2033 		if (br->br_chipid == BGE_ASICREV(chipid))
2034 			return (br);
2035 	}
2036 
2037 	return (NULL);
2038 }
2039 
2040 static const struct bge_product {
2041 	pci_vendor_id_t		bp_vendor;
2042 	pci_product_id_t	bp_product;
2043 	const char		*bp_name;
2044 } bge_products[] = {
2045 	/*
2046 	 * The BCM5700 documentation seems to indicate that the hardware
2047 	 * still has the Alteon vendor ID burned into it, though it
2048 	 * should always be overridden by the value in the EEPROM.  We'll
2049 	 * check for it anyway.
2050 	 */
2051 	{ PCI_VENDOR_ALTEON,
2052 	  PCI_PRODUCT_ALTEON_BCM5700,
2053 	  "Broadcom BCM5700 Gigabit Ethernet",
2054 	  },
2055 	{ PCI_VENDOR_ALTEON,
2056 	  PCI_PRODUCT_ALTEON_BCM5701,
2057 	  "Broadcom BCM5701 Gigabit Ethernet",
2058 	  },
2059 
2060 	{ PCI_VENDOR_ALTIMA,
2061 	  PCI_PRODUCT_ALTIMA_AC1000,
2062 	  "Altima AC1000 Gigabit Ethernet",
2063 	  },
2064 	{ PCI_VENDOR_ALTIMA,
2065 	  PCI_PRODUCT_ALTIMA_AC1001,
2066 	  "Altima AC1001 Gigabit Ethernet",
2067 	   },
2068 	{ PCI_VENDOR_ALTIMA,
2069 	  PCI_PRODUCT_ALTIMA_AC9100,
2070 	  "Altima AC9100 Gigabit Ethernet",
2071 	  },
2072 
2073 	{ PCI_VENDOR_BROADCOM,
2074 	  PCI_PRODUCT_BROADCOM_BCM5700,
2075 	  "Broadcom BCM5700 Gigabit Ethernet",
2076 	  },
2077 	{ PCI_VENDOR_BROADCOM,
2078 	  PCI_PRODUCT_BROADCOM_BCM5701,
2079 	  "Broadcom BCM5701 Gigabit Ethernet",
2080 	  },
2081 	{ PCI_VENDOR_BROADCOM,
2082 	  PCI_PRODUCT_BROADCOM_BCM5702,
2083 	  "Broadcom BCM5702 Gigabit Ethernet",
2084 	  },
2085 	{ PCI_VENDOR_BROADCOM,
2086 	  PCI_PRODUCT_BROADCOM_BCM5702X,
2087 	  "Broadcom BCM5702X Gigabit Ethernet" },
2088 
2089 	{ PCI_VENDOR_BROADCOM,
2090 	  PCI_PRODUCT_BROADCOM_BCM5703,
2091 	  "Broadcom BCM5703 Gigabit Ethernet",
2092 	  },
2093 	{ PCI_VENDOR_BROADCOM,
2094 	  PCI_PRODUCT_BROADCOM_BCM5703X,
2095 	  "Broadcom BCM5703X Gigabit Ethernet",
2096 	  },
2097 	{ PCI_VENDOR_BROADCOM,
2098 	  PCI_PRODUCT_BROADCOM_BCM5703A3,
2099 	  "Broadcom BCM5703A3 Gigabit Ethernet",
2100 	  },
2101 
2102    	{ PCI_VENDOR_BROADCOM,
2103 	  PCI_PRODUCT_BROADCOM_BCM5704C,
2104 	  "Broadcom BCM5704C Dual Gigabit Ethernet",
2105 	  },
2106    	{ PCI_VENDOR_BROADCOM,
2107 	  PCI_PRODUCT_BROADCOM_BCM5704S,
2108 	  "Broadcom BCM5704S Dual Gigabit Ethernet",
2109 	  },
2110 
2111    	{ PCI_VENDOR_BROADCOM,
2112 	  PCI_PRODUCT_BROADCOM_BCM5705,
2113 	  "Broadcom BCM5705 Gigabit Ethernet",
2114 	  },
2115    	{ PCI_VENDOR_BROADCOM,
2116 	  PCI_PRODUCT_BROADCOM_BCM5705K,
2117 	  "Broadcom BCM5705K Gigabit Ethernet",
2118 	  },
2119    	{ PCI_VENDOR_BROADCOM,
2120 	  PCI_PRODUCT_BROADCOM_BCM5705_ALT,
2121 	  "Broadcom BCM5705 Gigabit Ethernet",
2122 	  },
2123    	{ PCI_VENDOR_BROADCOM,
2124 	  PCI_PRODUCT_BROADCOM_BCM5705M,
2125 	  "Broadcom BCM5705M Gigabit Ethernet",
2126 	  },
2127 
2128 	{ PCI_VENDOR_BROADCOM,
2129 	  PCI_PRODUCT_BROADCOM_BCM5714,
2130 	  "Broadcom BCM5714/5715 Gigabit Ethernet",
2131 	  },
2132 	{ PCI_VENDOR_BROADCOM,
2133 	  PCI_PRODUCT_BROADCOM_BCM5789,
2134 	  "Broadcom BCM5789 Gigabit Ethernet",
2135 	  },
2136 
2137 	{ PCI_VENDOR_BROADCOM,
2138 	  PCI_PRODUCT_BROADCOM_BCM5721,
2139 	  "Broadcom BCM5721 Gigabit Ethernet",
2140 	  },
2141 
2142 	{ PCI_VENDOR_BROADCOM,
2143 	  PCI_PRODUCT_BROADCOM_BCM5750,
2144 	  "Broadcom BCM5750 Gigabit Ethernet",
2145 	  },
2146 
2147 	{ PCI_VENDOR_BROADCOM,
2148 	  PCI_PRODUCT_BROADCOM_BCM5750M,
2149 	  "Broadcom BCM5750M Gigabit Ethernet",
2150 	  },
2151 
2152 	{ PCI_VENDOR_BROADCOM,
2153 	  PCI_PRODUCT_BROADCOM_BCM5751,
2154 	  "Broadcom BCM5751 Gigabit Ethernet",
2155 	  },
2156 
2157 	{ PCI_VENDOR_BROADCOM,
2158 	  PCI_PRODUCT_BROADCOM_BCM5751M,
2159 	  "Broadcom BCM5751M Gigabit Ethernet",
2160 	  },
2161 
2162 	{ PCI_VENDOR_BROADCOM,
2163 	  PCI_PRODUCT_BROADCOM_BCM5752,
2164 	  "Broadcom BCM5752 Gigabit Ethernet",
2165 	  },
2166 
2167    	{ PCI_VENDOR_BROADCOM,
2168 	  PCI_PRODUCT_BROADCOM_BCM5782,
2169 	  "Broadcom BCM5782 Gigabit Ethernet",
2170 	  },
2171    	{ PCI_VENDOR_BROADCOM,
2172 	  PCI_PRODUCT_BROADCOM_BCM5788,
2173 	  "Broadcom BCM5788 Gigabit Ethernet",
2174 	  },
2175    	{ PCI_VENDOR_BROADCOM,
2176 	  PCI_PRODUCT_BROADCOM_BCM5789,
2177 	  "Broadcom BCM5789 Gigabit Ethernet",
2178 	  },
2179 
2180    	{ PCI_VENDOR_BROADCOM,
2181 	  PCI_PRODUCT_BROADCOM_BCM5901,
2182 	  "Broadcom BCM5901 Fast Ethernet",
2183 	  },
2184    	{ PCI_VENDOR_BROADCOM,
2185 	  PCI_PRODUCT_BROADCOM_BCM5901A2,
2186 	  "Broadcom BCM5901A2 Fast Ethernet",
2187 	  },
2188 
2189 	{ PCI_VENDOR_SCHNEIDERKOCH,
2190 	  PCI_PRODUCT_SCHNEIDERKOCH_SK_9DX1,
2191 	  "SysKonnect SK-9Dx1 Gigabit Ethernet",
2192 	  },
2193 
2194 	{ PCI_VENDOR_3COM,
2195 	  PCI_PRODUCT_3COM_3C996,
2196 	  "3Com 3c996 Gigabit Ethernet",
2197 	  },
2198 
2199 	{ 0,
2200 	  0,
2201 	  NULL },
2202 };
2203 
2204 static const struct bge_product *
2205 bge_lookup(const struct pci_attach_args *pa)
2206 {
2207 	const struct bge_product *bp;
2208 
2209 	for (bp = bge_products; bp->bp_name != NULL; bp++) {
2210 		if (PCI_VENDOR(pa->pa_id) == bp->bp_vendor &&
2211 		    PCI_PRODUCT(pa->pa_id) == bp->bp_product)
2212 			return (bp);
2213 	}
2214 
2215 	return (NULL);
2216 }
2217 
2218 static int
2219 bge_setpowerstate(struct bge_softc *sc, int powerlevel)
2220 {
2221 #ifdef NOTYET
2222 	u_int32_t pm_ctl = 0;
2223 
2224 	/* XXX FIXME: make sure indirect accesses enabled? */
2225 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_MISC_CTL, 4);
2226 	pm_ctl |= BGE_PCIMISCCTL_INDIRECT_ACCESS;
2227 	pci_write_config(sc->bge_dev, BGE_PCI_MISC_CTL, pm_ctl, 4);
2228 
2229 	/* clear the PME_assert bit and power state bits, enable PME */
2230 	pm_ctl = pci_conf_read(sc->bge_dev, BGE_PCI_PWRMGMT_CMD, 2);
2231 	pm_ctl &= ~PCIM_PSTAT_DMASK;
2232 	pm_ctl |= (1 << 8);
2233 
2234 	if (powerlevel == 0) {
2235 		pm_ctl |= PCIM_PSTAT_D0;
2236 		pci_write_config(sc->bge_dev, BGE_PCI_PWRMGMT_CMD,
2237 		    pm_ctl, 2);
2238 		DELAY(10000);
2239 		CSR_WRITE_4(sc, BGE_MISC_LOCAL_CTL, sc->bge_local_ctrl_reg);
2240 		DELAY(10000);
2241 
2242 #ifdef NOTYET
2243 		/* XXX FIXME: write 0x02 to phy aux_Ctrl reg */
2244 		bge_miibus_writereg(sc->bge_dev, 1, 0x18, 0x02);
2245 #endif
2246 		DELAY(40); DELAY(40); DELAY(40);
2247 		DELAY(10000);	/* above not quite adequate on 5700 */
2248 		return 0;
2249 	}
2250 
2251 
2252 	/*
2253 	 * Entering ACPI power states D1-D3 is achieved by wiggling
2254 	 * GMII gpio pins. Example code assumes all hardware vendors
2255 	 * followed Broadom's sample pcb layout. Until we verify that
2256 	 * for all supported OEM cards, states D1-D3 are  unsupported.
2257 	 */
2258 	printf("%s: power state %d unimplemented; check GPIO pins\n",
2259 	       sc->bge_dev.dv_xname, powerlevel);
2260 #endif
2261 	return EOPNOTSUPP;
2262 }
2263 
2264 
2265 /*
2266  * Probe for a Broadcom chip. Check the PCI vendor and device IDs
2267  * against our list and return its name if we find a match. Note
2268  * that since the Broadcom controller contains VPD support, we
2269  * can get the device name string from the controller itself instead
2270  * of the compiled-in string. This is a little slow, but it guarantees
2271  * we'll always announce the right product name.
2272  */
2273 static int
2274 bge_probe(device_t parent, cfdata_t match, void *aux)
2275 {
2276 	struct pci_attach_args *pa = (struct pci_attach_args *)aux;
2277 
2278 	if (bge_lookup(pa) != NULL)
2279 		return (1);
2280 
2281 	return (0);
2282 }
2283 
2284 static void
2285 bge_attach(device_t parent, device_t self, void *aux)
2286 {
2287 	struct bge_softc	*sc = (struct bge_softc *)self;
2288 	struct pci_attach_args	*pa = aux;
2289 	const struct bge_product *bp;
2290 	const struct bge_revision *br;
2291 	pci_chipset_tag_t	pc = pa->pa_pc;
2292 	pci_intr_handle_t	ih;
2293 	const char		*intrstr = NULL;
2294 	bus_dma_segment_t	seg;
2295 	int			rseg;
2296 	u_int32_t		hwcfg = 0;
2297 	u_int32_t		mac_addr = 0;
2298 	u_int32_t		command;
2299 	struct ifnet		*ifp;
2300 	caddr_t			kva;
2301 	u_char			eaddr[ETHER_ADDR_LEN];
2302 	pcireg_t		memtype;
2303 	bus_addr_t		memaddr;
2304 	bus_size_t		memsize;
2305 	u_int32_t		pm_ctl;
2306 
2307 	bp = bge_lookup(pa);
2308 	KASSERT(bp != NULL);
2309 
2310 	sc->bge_pa = *pa;
2311 
2312 	aprint_naive(": Ethernet controller\n");
2313 	aprint_normal(": %s\n", bp->bp_name);
2314 
2315 	/*
2316 	 * Map control/status registers.
2317 	 */
2318 	DPRINTFN(5, ("Map control/status regs\n"));
2319 	command = pci_conf_read(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG);
2320 	command |= PCI_COMMAND_MEM_ENABLE | PCI_COMMAND_MASTER_ENABLE;
2321 	pci_conf_write(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG, command);
2322 	command = pci_conf_read(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG);
2323 
2324 	if (!(command & PCI_COMMAND_MEM_ENABLE)) {
2325 		aprint_error("%s: failed to enable memory mapping!\n",
2326 		    sc->bge_dev.dv_xname);
2327 		return;
2328 	}
2329 
2330 	DPRINTFN(5, ("pci_mem_find\n"));
2331 	memtype = pci_mapreg_type(pa->pa_pc, pa->pa_tag, BGE_PCI_BAR0);
2332  	switch (memtype) {
2333 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_32BIT:
2334 	case PCI_MAPREG_TYPE_MEM | PCI_MAPREG_MEM_TYPE_64BIT:
2335 		if (pci_mapreg_map(pa, BGE_PCI_BAR0,
2336 		    memtype, 0, &sc->bge_btag, &sc->bge_bhandle,
2337 		    &memaddr, &memsize) == 0)
2338 			break;
2339 	default:
2340 		aprint_error("%s: can't find mem space\n",
2341 		    sc->bge_dev.dv_xname);
2342 		return;
2343 	}
2344 
2345 	DPRINTFN(5, ("pci_intr_map\n"));
2346 	if (pci_intr_map(pa, &ih)) {
2347 		aprint_error("%s: couldn't map interrupt\n",
2348 		    sc->bge_dev.dv_xname);
2349 		return;
2350 	}
2351 
2352 	DPRINTFN(5, ("pci_intr_string\n"));
2353 	intrstr = pci_intr_string(pc, ih);
2354 
2355 	DPRINTFN(5, ("pci_intr_establish\n"));
2356 	sc->bge_intrhand = pci_intr_establish(pc, ih, IPL_NET, bge_intr, sc);
2357 
2358 	if (sc->bge_intrhand == NULL) {
2359 		aprint_error("%s: couldn't establish interrupt",
2360 		    sc->bge_dev.dv_xname);
2361 		if (intrstr != NULL)
2362 			aprint_normal(" at %s", intrstr);
2363 		aprint_normal("\n");
2364 		return;
2365 	}
2366 	aprint_normal("%s: interrupting at %s\n",
2367 	    sc->bge_dev.dv_xname, intrstr);
2368 
2369 	/*
2370 	 * Kludge for 5700 Bx bug: a hardware bug (PCIX byte enable?)
2371 	 * can clobber the chip's PCI config-space power control registers,
2372 	 * leaving the card in D3 powersave state.
2373 	 * We do not have memory-mapped registers in this state,
2374 	 * so force device into D0 state before starting initialization.
2375 	 */
2376 	pm_ctl = pci_conf_read(pc, pa->pa_tag, BGE_PCI_PWRMGMT_CMD);
2377 	pm_ctl &= ~(PCI_PWR_D0|PCI_PWR_D1|PCI_PWR_D2|PCI_PWR_D3);
2378 	pm_ctl |= (1 << 8) | PCI_PWR_D0 ; /* D0 state */
2379 	pci_conf_write(pc, pa->pa_tag, BGE_PCI_PWRMGMT_CMD, pm_ctl);
2380 	DELAY(1000);	/* 27 usec is allegedly sufficent */
2381 
2382 	/*
2383 	 * Save ASIC rev.  Look up any quirks associated with this
2384 	 * ASIC.
2385 	 */
2386 	sc->bge_chipid =
2387 	    pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL) &
2388 	    BGE_PCIMISCCTL_ASICREV;
2389 
2390 	/*
2391 	 * Detect PCI-Express devices
2392 	 * XXX: guessed from Linux/FreeBSD; no documentation
2393 	 */
2394 	if (BGE_ASICREV(sc->bge_chipid) == BGE_ASICREV_BCM5750 &&
2395 	    pci_get_capability(pa->pa_pc, pa->pa_tag, PCI_CAP_PCIEXPRESS,
2396 	    NULL, NULL) != 0)
2397 		sc->bge_pcie = 1;
2398 	else
2399 		sc->bge_pcie = 0;
2400 
2401 	/* Try to reset the chip. */
2402 	DPRINTFN(5, ("bge_reset\n"));
2403 	bge_reset(sc);
2404 
2405 	if (bge_chipinit(sc)) {
2406 		aprint_error("%s: chip initialization failed\n",
2407 		    sc->bge_dev.dv_xname);
2408 		bge_release_resources(sc);
2409 		return;
2410 	}
2411 
2412 	/*
2413 	 * Get station address from the EEPROM.
2414 	 */
2415 	mac_addr = bge_readmem_ind(sc, 0x0c14);
2416 	if ((mac_addr >> 16) == 0x484b) {
2417 		eaddr[0] = (u_char)(mac_addr >> 8);
2418 		eaddr[1] = (u_char)(mac_addr >> 0);
2419 		mac_addr = bge_readmem_ind(sc, 0x0c18);
2420 		eaddr[2] = (u_char)(mac_addr >> 24);
2421 		eaddr[3] = (u_char)(mac_addr >> 16);
2422 		eaddr[4] = (u_char)(mac_addr >> 8);
2423 		eaddr[5] = (u_char)(mac_addr >> 0);
2424 	} else if (bge_read_eeprom(sc, (caddr_t)eaddr,
2425 	    BGE_EE_MAC_OFFSET + 2, ETHER_ADDR_LEN)) {
2426 		aprint_error("%s: failed to read station address\n",
2427 		    sc->bge_dev.dv_xname);
2428 		bge_release_resources(sc);
2429 		return;
2430 	}
2431 
2432 	br = bge_lookup_rev(sc->bge_chipid);
2433 	aprint_normal("%s: ", sc->bge_dev.dv_xname);
2434 
2435 	if (br == NULL) {
2436 		aprint_normal("unknown ASIC (0x%04x)", sc->bge_chipid >> 16);
2437 		sc->bge_quirks = 0;
2438 	} else {
2439 		aprint_normal("ASIC %s (0x%04x)",
2440 		    br->br_name, sc->bge_chipid >> 16);
2441 		sc->bge_quirks |= br->br_quirks;
2442 	}
2443 	aprint_normal(", Ethernet address %s\n", ether_sprintf(eaddr));
2444 
2445 	/* Allocate the general information block and ring buffers. */
2446 	if (pci_dma64_available(pa))
2447 		sc->bge_dmatag = pa->pa_dmat64;
2448 	else
2449 		sc->bge_dmatag = pa->pa_dmat;
2450 	DPRINTFN(5, ("bus_dmamem_alloc\n"));
2451 	if (bus_dmamem_alloc(sc->bge_dmatag, sizeof(struct bge_ring_data),
2452 			     PAGE_SIZE, 0, &seg, 1, &rseg, BUS_DMA_NOWAIT)) {
2453 		aprint_error("%s: can't alloc rx buffers\n",
2454 		    sc->bge_dev.dv_xname);
2455 		return;
2456 	}
2457 	DPRINTFN(5, ("bus_dmamem_map\n"));
2458 	if (bus_dmamem_map(sc->bge_dmatag, &seg, rseg,
2459 			   sizeof(struct bge_ring_data), &kva,
2460 			   BUS_DMA_NOWAIT)) {
2461 		aprint_error("%s: can't map DMA buffers (%d bytes)\n",
2462 		    sc->bge_dev.dv_xname, (int)sizeof(struct bge_ring_data));
2463 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2464 		return;
2465 	}
2466 	DPRINTFN(5, ("bus_dmamem_create\n"));
2467 	if (bus_dmamap_create(sc->bge_dmatag, sizeof(struct bge_ring_data), 1,
2468 	    sizeof(struct bge_ring_data), 0,
2469 	    BUS_DMA_NOWAIT, &sc->bge_ring_map)) {
2470 		aprint_error("%s: can't create DMA map\n",
2471 		    sc->bge_dev.dv_xname);
2472 		bus_dmamem_unmap(sc->bge_dmatag, kva,
2473 				 sizeof(struct bge_ring_data));
2474 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2475 		return;
2476 	}
2477 	DPRINTFN(5, ("bus_dmamem_load\n"));
2478 	if (bus_dmamap_load(sc->bge_dmatag, sc->bge_ring_map, kva,
2479 			    sizeof(struct bge_ring_data), NULL,
2480 			    BUS_DMA_NOWAIT)) {
2481 		bus_dmamap_destroy(sc->bge_dmatag, sc->bge_ring_map);
2482 		bus_dmamem_unmap(sc->bge_dmatag, kva,
2483 				 sizeof(struct bge_ring_data));
2484 		bus_dmamem_free(sc->bge_dmatag, &seg, rseg);
2485 		return;
2486 	}
2487 
2488 	DPRINTFN(5, ("bzero\n"));
2489 	sc->bge_rdata = (struct bge_ring_data *)kva;
2490 
2491 	memset(sc->bge_rdata, 0, sizeof(struct bge_ring_data));
2492 
2493 	/* Try to allocate memory for jumbo buffers. */
2494 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
2495 		if (bge_alloc_jumbo_mem(sc)) {
2496 			aprint_error("%s: jumbo buffer allocation failed\n",
2497 			    sc->bge_dev.dv_xname);
2498 		} else
2499 			sc->ethercom.ec_capabilities |= ETHERCAP_JUMBO_MTU;
2500 	}
2501 
2502 	/* Set default tuneable values. */
2503 	sc->bge_stat_ticks = BGE_TICKS_PER_SEC;
2504 	sc->bge_rx_coal_ticks = 150;
2505 	sc->bge_rx_max_coal_bds = 64;
2506 #ifdef ORIG_WPAUL_VALUES
2507 	sc->bge_tx_coal_ticks = 150;
2508 	sc->bge_tx_max_coal_bds = 128;
2509 #else
2510 	sc->bge_tx_coal_ticks = 300;
2511 	sc->bge_tx_max_coal_bds = 400;
2512 #endif
2513 	if (sc->bge_quirks & BGE_QUIRK_5705_CORE) {
2514 		sc->bge_tx_coal_ticks = (12 * 5);
2515 		sc->bge_rx_max_coal_bds = (12 * 5);
2516 			aprint_error("%s: setting short Tx thresholds\n",
2517 			    sc->bge_dev.dv_xname);
2518 	}
2519 
2520 	/* Set up ifnet structure */
2521 	ifp = &sc->ethercom.ec_if;
2522 	ifp->if_softc = sc;
2523 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
2524 	ifp->if_ioctl = bge_ioctl;
2525 	ifp->if_start = bge_start;
2526 	ifp->if_init = bge_init;
2527 	ifp->if_watchdog = bge_watchdog;
2528 	IFQ_SET_MAXLEN(&ifp->if_snd, max(BGE_TX_RING_CNT - 1, IFQ_MAXLEN));
2529 	IFQ_SET_READY(&ifp->if_snd);
2530 	DPRINTFN(5, ("bcopy\n"));
2531 	strcpy(ifp->if_xname, sc->bge_dev.dv_xname);
2532 
2533 	if ((sc->bge_quirks & BGE_QUIRK_CSUM_BROKEN) == 0)
2534 		sc->ethercom.ec_if.if_capabilities |=
2535 		    IFCAP_CSUM_IPv4_Tx | IFCAP_CSUM_IPv4_Rx |
2536 		    IFCAP_CSUM_TCPv4_Tx | IFCAP_CSUM_TCPv4_Rx |
2537 		    IFCAP_CSUM_UDPv4_Tx | IFCAP_CSUM_UDPv4_Rx;
2538 	sc->ethercom.ec_capabilities |=
2539 	    ETHERCAP_VLAN_HWTAGGING | ETHERCAP_VLAN_MTU;
2540 
2541 	if (sc->bge_pcie)
2542 		sc->ethercom.ec_if.if_capabilities |= IFCAP_TSOv4;
2543 
2544 	/*
2545 	 * Do MII setup.
2546 	 */
2547 	DPRINTFN(5, ("mii setup\n"));
2548 	sc->bge_mii.mii_ifp = ifp;
2549 	sc->bge_mii.mii_readreg = bge_miibus_readreg;
2550 	sc->bge_mii.mii_writereg = bge_miibus_writereg;
2551 	sc->bge_mii.mii_statchg = bge_miibus_statchg;
2552 
2553 	/*
2554 	 * Figure out what sort of media we have by checking the
2555 	 * hardware config word in the first 32k of NIC internal memory,
2556 	 * or fall back to the config word in the EEPROM. Note: on some BCM5700
2557 	 * cards, this value appears to be unset. If that's the
2558 	 * case, we have to rely on identifying the NIC by its PCI
2559 	 * subsystem ID, as we do below for the SysKonnect SK-9D41.
2560 	 */
2561 	if (bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_SIG) == BGE_MAGIC_NUMBER) {
2562 		hwcfg = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM_NICCFG);
2563 	} else {
2564 		bge_read_eeprom(sc, (caddr_t)&hwcfg,
2565 		    BGE_EE_HWCFG_OFFSET, sizeof(hwcfg));
2566 		hwcfg = be32toh(hwcfg);
2567 	}
2568 	if ((hwcfg & BGE_HWCFG_MEDIA) == BGE_MEDIA_FIBER)
2569 		sc->bge_tbi = 1;
2570 
2571 	/* The SysKonnect SK-9D41 is a 1000baseSX card. */
2572 	if ((pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_SUBSYS) >> 16) ==
2573 	    SK_SUBSYSID_9D41)
2574 		sc->bge_tbi = 1;
2575 
2576 	if (sc->bge_tbi) {
2577 		ifmedia_init(&sc->bge_ifmedia, IFM_IMASK, bge_ifmedia_upd,
2578 		    bge_ifmedia_sts);
2579 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX, 0, NULL);
2580 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_1000_SX|IFM_FDX,
2581 			    0, NULL);
2582 		ifmedia_add(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO, 0, NULL);
2583 		ifmedia_set(&sc->bge_ifmedia, IFM_ETHER|IFM_AUTO);
2584 	} else {
2585 		/*
2586 		 * Do transceiver setup.
2587 		 */
2588 		ifmedia_init(&sc->bge_mii.mii_media, 0, bge_ifmedia_upd,
2589 			     bge_ifmedia_sts);
2590 		mii_attach(&sc->bge_dev, &sc->bge_mii, 0xffffffff,
2591 			   MII_PHY_ANY, MII_OFFSET_ANY,
2592 			   MIIF_FORCEANEG|MIIF_DOPAUSE);
2593 
2594 		if (LIST_FIRST(&sc->bge_mii.mii_phys) == NULL) {
2595 			printf("%s: no PHY found!\n", sc->bge_dev.dv_xname);
2596 			ifmedia_add(&sc->bge_mii.mii_media,
2597 				    IFM_ETHER|IFM_MANUAL, 0, NULL);
2598 			ifmedia_set(&sc->bge_mii.mii_media,
2599 				    IFM_ETHER|IFM_MANUAL);
2600 		} else
2601 			ifmedia_set(&sc->bge_mii.mii_media,
2602 				    IFM_ETHER|IFM_AUTO);
2603 	}
2604 
2605 	/*
2606 	 * When using the BCM5701 in PCI-X mode, data corruption has
2607 	 * been observed in the first few bytes of some received packets.
2608 	 * Aligning the packet buffer in memory eliminates the corruption.
2609 	 * Unfortunately, this misaligns the packet payloads.  On platforms
2610 	 * which do not support unaligned accesses, we will realign the
2611 	 * payloads by copying the received packets.
2612 	 */
2613 	if (sc->bge_quirks & BGE_QUIRK_PCIX_DMA_ALIGN_BUG) {
2614 		/* If in PCI-X mode, work around the alignment bug. */
2615 		if ((pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_PCISTATE) &
2616                     (BGE_PCISTATE_PCI_BUSMODE | BGE_PCISTATE_PCI_BUSSPEED)) ==
2617                          BGE_PCISTATE_PCI_BUSSPEED)
2618 		sc->bge_rx_alignment_bug = 1;
2619         }
2620 
2621 	/*
2622 	 * Call MI attach routine.
2623 	 */
2624 	DPRINTFN(5, ("if_attach\n"));
2625 	if_attach(ifp);
2626 	DPRINTFN(5, ("ether_ifattach\n"));
2627 	ether_ifattach(ifp, eaddr);
2628 #ifdef BGE_EVENT_COUNTERS
2629 	/*
2630 	 * Attach event counters.
2631 	 */
2632 	evcnt_attach_dynamic(&sc->bge_ev_intr, EVCNT_TYPE_INTR,
2633 	    NULL, sc->bge_dev.dv_xname, "intr");
2634 	evcnt_attach_dynamic(&sc->bge_ev_tx_xoff, EVCNT_TYPE_MISC,
2635 	    NULL, sc->bge_dev.dv_xname, "tx_xoff");
2636 	evcnt_attach_dynamic(&sc->bge_ev_tx_xon, EVCNT_TYPE_MISC,
2637 	    NULL, sc->bge_dev.dv_xname, "tx_xon");
2638 	evcnt_attach_dynamic(&sc->bge_ev_rx_xoff, EVCNT_TYPE_MISC,
2639 	    NULL, sc->bge_dev.dv_xname, "rx_xoff");
2640 	evcnt_attach_dynamic(&sc->bge_ev_rx_xon, EVCNT_TYPE_MISC,
2641 	    NULL, sc->bge_dev.dv_xname, "rx_xon");
2642 	evcnt_attach_dynamic(&sc->bge_ev_rx_macctl, EVCNT_TYPE_MISC,
2643 	    NULL, sc->bge_dev.dv_xname, "rx_macctl");
2644 	evcnt_attach_dynamic(&sc->bge_ev_xoffentered, EVCNT_TYPE_MISC,
2645 	    NULL, sc->bge_dev.dv_xname, "xoffentered");
2646 #endif /* BGE_EVENT_COUNTERS */
2647 	DPRINTFN(5, ("callout_init\n"));
2648 	callout_init(&sc->bge_timeout);
2649 
2650 	sc->bge_powerhook = powerhook_establish(bge_powerhook, sc);
2651 	if (sc->bge_powerhook == NULL)
2652 		printf("%s: WARNING: unable to establish PCI power hook\n",
2653 		    sc->bge_dev.dv_xname);
2654 }
2655 
2656 static void
2657 bge_release_resources(struct bge_softc *sc)
2658 {
2659 	if (sc->bge_vpd_prodname != NULL)
2660 		free(sc->bge_vpd_prodname, M_DEVBUF);
2661 
2662 	if (sc->bge_vpd_readonly != NULL)
2663 		free(sc->bge_vpd_readonly, M_DEVBUF);
2664 }
2665 
2666 static void
2667 bge_reset(struct bge_softc *sc)
2668 {
2669 	struct pci_attach_args *pa = &sc->bge_pa;
2670 	u_int32_t cachesize, command, pcistate, new_pcistate;
2671 	int i, val;
2672 
2673 	/* Save some important PCI state. */
2674 	cachesize = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CACHESZ);
2675 	command = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD);
2676 	pcistate = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_PCISTATE);
2677 
2678 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL,
2679 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
2680 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
2681 
2682 	val = BGE_MISCCFG_RESET_CORE_CLOCKS | (65<<1);
2683 	/*
2684 	 * XXX: from FreeBSD/Linux; no documentation
2685 	 */
2686 	if (sc->bge_pcie) {
2687 		if (CSR_READ_4(sc, BGE_PCIE_CTL1) == 0x60)
2688 			CSR_WRITE_4(sc, BGE_PCIE_CTL1, 0x20);
2689 		if (sc->bge_chipid != BGE_CHIPID_BCM5750_A0) {
2690 			/* No idea what that actually means */
2691 			CSR_WRITE_4(sc, BGE_MISC_CFG, 1 << 29);
2692 			val |= (1<<29);
2693 		}
2694 	}
2695 
2696 	/* Issue global reset */
2697 	bge_writereg_ind(sc, BGE_MISC_CFG, val);
2698 
2699 	DELAY(1000);
2700 
2701 	/*
2702 	 * XXX: from FreeBSD/Linux; no documentation
2703 	 */
2704 	if (sc->bge_pcie) {
2705 		if (sc->bge_chipid == BGE_CHIPID_BCM5750_A0) {
2706 			pcireg_t reg;
2707 
2708 			DELAY(500000);
2709 			/* XXX: Magic Numbers */
2710 			reg = pci_conf_read(pa->pa_pc, pa->pa_tag, BGE_PCI_UNKNOWN0);
2711 			pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_UNKNOWN0,
2712 			    reg | (1 << 15));
2713 		}
2714 		/*
2715 		 * XXX: Magic Numbers.
2716 		 * Sets maximal PCI-e payload and clears any PCI-e errors.
2717 		 * Should be replaced with references to PCI config-space
2718 		 * capability block for PCI-Express.
2719 		 */
2720 		pci_conf_write(pa->pa_pc, pa->pa_tag,
2721 		    BGE_PCI_CONF_DEV_CTRL, 0xf5000);
2722 
2723 	}
2724 
2725 	/* Reset some of the PCI state that got zapped by reset */
2726 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_MISC_CTL,
2727 	    BGE_PCIMISCCTL_INDIRECT_ACCESS|BGE_PCIMISCCTL_MASK_PCI_INTR|
2728 	    BGE_HIF_SWAP_OPTIONS|BGE_PCIMISCCTL_PCISTATE_RW);
2729 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_CMD, command);
2730 	pci_conf_write(pa->pa_pc, pa->pa_tag, BGE_PCI_CACHESZ, cachesize);
2731 	bge_writereg_ind(sc, BGE_MISC_CFG, (65 << 1));
2732 
2733 	/* Enable memory arbiter. */
2734 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
2735 		uint32_t marbmode = 0;
2736 		if (BGE_IS_5714_FAMILY(sc)) {
2737 			marbmode = CSR_READ_4(sc, BGE_MARB_MODE);
2738 		}
2739  		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE | marbmode);
2740 	}
2741 
2742 	/*
2743 	 * Prevent PXE restart: write a magic number to the
2744 	 * general communications memory at 0xB50.
2745 	 */
2746 	bge_writemem_ind(sc, BGE_SOFTWARE_GENCOMM, BGE_MAGIC_NUMBER);
2747 
2748 	/*
2749 	 * Poll the value location we just wrote until
2750 	 * we see the 1's complement of the magic number.
2751 	 * This indicates that the firmware initialization
2752 	 * is complete.
2753 	 */
2754 	for (i = 0; i < BGE_TIMEOUT; i++) {
2755 		val = bge_readmem_ind(sc, BGE_SOFTWARE_GENCOMM);
2756 		if (val == ~BGE_MAGIC_NUMBER)
2757 			break;
2758 		DELAY(1000);
2759 	}
2760 
2761 	if (i >= BGE_TIMEOUT) {
2762 		printf("%s: firmware handshake timed out, val = %x\n",
2763 		    sc->bge_dev.dv_xname, val);
2764 		/*
2765 		 * XXX: occasionally fired on bcm5721, but without
2766 		 * apparent harm.  For now, keep going if we timeout
2767 		 * against PCI-E devices.
2768 		 */
2769 		 if (!sc->bge_pcie)
2770 		  return;
2771 	}
2772 
2773 	/*
2774 	 * XXX Wait for the value of the PCISTATE register to
2775 	 * return to its original pre-reset state. This is a
2776 	 * fairly good indicator of reset completion. If we don't
2777 	 * wait for the reset to fully complete, trying to read
2778 	 * from the device's non-PCI registers may yield garbage
2779 	 * results.
2780 	 */
2781 	for (i = 0; i < BGE_TIMEOUT; i++) {
2782 		new_pcistate = pci_conf_read(pa->pa_pc, pa->pa_tag,
2783 		    BGE_PCI_PCISTATE);
2784 		if ((new_pcistate & ~BGE_PCISTATE_RESERVED) ==
2785 		    (pcistate & ~BGE_PCISTATE_RESERVED))
2786 			break;
2787 		DELAY(10);
2788 	}
2789 	if ((new_pcistate & ~BGE_PCISTATE_RESERVED) !=
2790 	    (pcistate & ~BGE_PCISTATE_RESERVED)) {
2791 		printf("%s: pcistate failed to revert\n",
2792 		    sc->bge_dev.dv_xname);
2793 	}
2794 
2795 	/* XXX: from FreeBSD/Linux; no documentation */
2796 	if (sc->bge_pcie && sc->bge_chipid != BGE_CHIPID_BCM5750_A0)
2797 		CSR_WRITE_4(sc, BGE_PCIE_CTL0, CSR_READ_4(sc, BGE_PCIE_CTL0) | (1<<25));
2798 
2799 	/* Enable memory arbiter. */
2800 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
2801 		uint32_t marbmode = 0;
2802 		if (BGE_IS_5714_FAMILY(sc)) {
2803 			marbmode = CSR_READ_4(sc, BGE_MARB_MODE);
2804 		}
2805  		CSR_WRITE_4(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE | marbmode);
2806 	}
2807 
2808 	/* Fix up byte swapping */
2809 	CSR_WRITE_4(sc, BGE_MODE_CTL, BGE_DMA_SWAP_OPTIONS);
2810 
2811 	CSR_WRITE_4(sc, BGE_MAC_MODE, 0);
2812 
2813 	DELAY(10000);
2814 }
2815 
2816 /*
2817  * Frame reception handling. This is called if there's a frame
2818  * on the receive return list.
2819  *
2820  * Note: we have to be able to handle two possibilities here:
2821  * 1) the frame is from the jumbo recieve ring
2822  * 2) the frame is from the standard receive ring
2823  */
2824 
2825 static void
2826 bge_rxeof(struct bge_softc *sc)
2827 {
2828 	struct ifnet *ifp;
2829 	int stdcnt = 0, jumbocnt = 0;
2830 	bus_dmamap_t dmamap;
2831 	bus_addr_t offset, toff;
2832 	bus_size_t tlen;
2833 	int tosync;
2834 
2835 	ifp = &sc->ethercom.ec_if;
2836 
2837 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2838 	    offsetof(struct bge_ring_data, bge_status_block),
2839 	    sizeof (struct bge_status_block),
2840 	    BUS_DMASYNC_POSTREAD);
2841 
2842 	offset = offsetof(struct bge_ring_data, bge_rx_return_ring);
2843 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx -
2844 	    sc->bge_rx_saved_considx;
2845 
2846 	toff = offset + (sc->bge_rx_saved_considx * sizeof (struct bge_rx_bd));
2847 
2848 	if (tosync < 0) {
2849 		tlen = (sc->bge_return_ring_cnt - sc->bge_rx_saved_considx) *
2850 		    sizeof (struct bge_rx_bd);
2851 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2852 		    toff, tlen, BUS_DMASYNC_POSTREAD);
2853 		tosync = -tosync;
2854 	}
2855 
2856 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2857 	    offset, tosync * sizeof (struct bge_rx_bd),
2858 	    BUS_DMASYNC_POSTREAD);
2859 
2860 	while(sc->bge_rx_saved_considx !=
2861 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_rx_prod_idx) {
2862 		struct bge_rx_bd	*cur_rx;
2863 		u_int32_t		rxidx;
2864 		struct mbuf		*m = NULL;
2865 
2866 		cur_rx = &sc->bge_rdata->
2867 			bge_rx_return_ring[sc->bge_rx_saved_considx];
2868 
2869 		rxidx = cur_rx->bge_idx;
2870 		BGE_INC(sc->bge_rx_saved_considx, sc->bge_return_ring_cnt);
2871 
2872 		if (cur_rx->bge_flags & BGE_RXBDFLAG_JUMBO_RING) {
2873 			BGE_INC(sc->bge_jumbo, BGE_JUMBO_RX_RING_CNT);
2874 			m = sc->bge_cdata.bge_rx_jumbo_chain[rxidx];
2875 			sc->bge_cdata.bge_rx_jumbo_chain[rxidx] = NULL;
2876 			jumbocnt++;
2877 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
2878 				ifp->if_ierrors++;
2879 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
2880 				continue;
2881 			}
2882 			if (bge_newbuf_jumbo(sc, sc->bge_jumbo,
2883 					     NULL)== ENOBUFS) {
2884 				ifp->if_ierrors++;
2885 				bge_newbuf_jumbo(sc, sc->bge_jumbo, m);
2886 				continue;
2887 			}
2888 		} else {
2889 			BGE_INC(sc->bge_std, BGE_STD_RX_RING_CNT);
2890 			m = sc->bge_cdata.bge_rx_std_chain[rxidx];
2891 			sc->bge_cdata.bge_rx_std_chain[rxidx] = NULL;
2892 			stdcnt++;
2893 			dmamap = sc->bge_cdata.bge_rx_std_map[rxidx];
2894 			sc->bge_cdata.bge_rx_std_map[rxidx] = 0;
2895 			if (cur_rx->bge_flags & BGE_RXBDFLAG_ERROR) {
2896 				ifp->if_ierrors++;
2897 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
2898 				continue;
2899 			}
2900 			if (bge_newbuf_std(sc, sc->bge_std,
2901 			    NULL, dmamap) == ENOBUFS) {
2902 				ifp->if_ierrors++;
2903 				bge_newbuf_std(sc, sc->bge_std, m, dmamap);
2904 				continue;
2905 			}
2906 		}
2907 
2908 		ifp->if_ipackets++;
2909 #ifndef __NO_STRICT_ALIGNMENT
2910                 /*
2911                  * XXX: if the 5701 PCIX-Rx-DMA workaround is in effect,
2912                  * the Rx buffer has the layer-2 header unaligned.
2913                  * If our CPU requires alignment, re-align by copying.
2914                  */
2915 		if (sc->bge_rx_alignment_bug) {
2916 			memmove(mtod(m, caddr_t) + ETHER_ALIGN, m->m_data,
2917                                 cur_rx->bge_len);
2918 			m->m_data += ETHER_ALIGN;
2919 		}
2920 #endif
2921 
2922 		m->m_pkthdr.len = m->m_len = cur_rx->bge_len - ETHER_CRC_LEN;
2923 		m->m_pkthdr.rcvif = ifp;
2924 
2925 #if NBPFILTER > 0
2926 		/*
2927 		 * Handle BPF listeners. Let the BPF user see the packet.
2928 		 */
2929 		if (ifp->if_bpf)
2930 			bpf_mtap(ifp->if_bpf, m);
2931 #endif
2932 
2933 		m->m_pkthdr.csum_flags = M_CSUM_IPv4;
2934 
2935 		if ((cur_rx->bge_ip_csum ^ 0xffff) != 0)
2936 			m->m_pkthdr.csum_flags |= M_CSUM_IPv4_BAD;
2937 		/*
2938 		 * Rx transport checksum-offload may also
2939 		 * have bugs with packets which, when transmitted,
2940 		 * were `runts' requiring padding.
2941 		 */
2942 		if (cur_rx->bge_flags & BGE_RXBDFLAG_TCP_UDP_CSUM &&
2943 		    (/* (sc->_bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||*/
2944 		     m->m_pkthdr.len >= ETHER_MIN_NOPAD)) {
2945 			m->m_pkthdr.csum_data =
2946 			    cur_rx->bge_tcp_udp_csum;
2947 			m->m_pkthdr.csum_flags |=
2948 			    (M_CSUM_TCPv4|M_CSUM_UDPv4|
2949 			     M_CSUM_DATA|M_CSUM_NO_PSEUDOHDR);
2950 		}
2951 
2952 		/*
2953 		 * If we received a packet with a vlan tag, pass it
2954 		 * to vlan_input() instead of ether_input().
2955 		 */
2956 		if (cur_rx->bge_flags & BGE_RXBDFLAG_VLAN_TAG)
2957 			VLAN_INPUT_TAG(ifp, m, cur_rx->bge_vlan_tag, continue);
2958 
2959 		(*ifp->if_input)(ifp, m);
2960 	}
2961 
2962 	CSR_WRITE_4(sc, BGE_MBX_RX_CONS0_LO, sc->bge_rx_saved_considx);
2963 	if (stdcnt)
2964 		CSR_WRITE_4(sc, BGE_MBX_RX_STD_PROD_LO, sc->bge_std);
2965 	if (jumbocnt)
2966 		CSR_WRITE_4(sc, BGE_MBX_RX_JUMBO_PROD_LO, sc->bge_jumbo);
2967 }
2968 
2969 static void
2970 bge_txeof(struct bge_softc *sc)
2971 {
2972 	struct bge_tx_bd *cur_tx = NULL;
2973 	struct ifnet *ifp;
2974 	struct txdmamap_pool_entry *dma;
2975 	bus_addr_t offset, toff;
2976 	bus_size_t tlen;
2977 	int tosync;
2978 	struct mbuf *m;
2979 
2980 	ifp = &sc->ethercom.ec_if;
2981 
2982 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2983 	    offsetof(struct bge_ring_data, bge_status_block),
2984 	    sizeof (struct bge_status_block),
2985 	    BUS_DMASYNC_POSTREAD);
2986 
2987 	offset = offsetof(struct bge_ring_data, bge_tx_ring);
2988 	tosync = sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx -
2989 	    sc->bge_tx_saved_considx;
2990 
2991 	toff = offset + (sc->bge_tx_saved_considx * sizeof (struct bge_tx_bd));
2992 
2993 	if (tosync < 0) {
2994 		tlen = (BGE_TX_RING_CNT - sc->bge_tx_saved_considx) *
2995 		    sizeof (struct bge_tx_bd);
2996 		bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
2997 		    toff, tlen, BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
2998 		tosync = -tosync;
2999 	}
3000 
3001 	bus_dmamap_sync(sc->bge_dmatag, sc->bge_ring_map,
3002 	    offset, tosync * sizeof (struct bge_tx_bd),
3003 	    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
3004 
3005 	/*
3006 	 * Go through our tx ring and free mbufs for those
3007 	 * frames that have been sent.
3008 	 */
3009 	while (sc->bge_tx_saved_considx !=
3010 	    sc->bge_rdata->bge_status_block.bge_idx[0].bge_tx_cons_idx) {
3011 		u_int32_t		idx = 0;
3012 
3013 		idx = sc->bge_tx_saved_considx;
3014 		cur_tx = &sc->bge_rdata->bge_tx_ring[idx];
3015 		if (cur_tx->bge_flags & BGE_TXBDFLAG_END)
3016 			ifp->if_opackets++;
3017 		m = sc->bge_cdata.bge_tx_chain[idx];
3018 		if (m != NULL) {
3019 			sc->bge_cdata.bge_tx_chain[idx] = NULL;
3020 			dma = sc->txdma[idx];
3021 			bus_dmamap_sync(sc->bge_dmatag, dma->dmamap, 0,
3022 			    dma->dmamap->dm_mapsize, BUS_DMASYNC_POSTWRITE);
3023 			bus_dmamap_unload(sc->bge_dmatag, dma->dmamap);
3024 			SLIST_INSERT_HEAD(&sc->txdma_list, dma, link);
3025 			sc->txdma[idx] = NULL;
3026 
3027 			m_freem(m);
3028 		}
3029 		sc->bge_txcnt--;
3030 		BGE_INC(sc->bge_tx_saved_considx, BGE_TX_RING_CNT);
3031 		ifp->if_timer = 0;
3032 	}
3033 
3034 	if (cur_tx != NULL)
3035 		ifp->if_flags &= ~IFF_OACTIVE;
3036 }
3037 
3038 static int
3039 bge_intr(void *xsc)
3040 {
3041 	struct bge_softc *sc;
3042 	struct ifnet *ifp;
3043 
3044 	sc = xsc;
3045 	ifp = &sc->ethercom.ec_if;
3046 
3047 #ifdef notdef
3048 	/* Avoid this for now -- checking this register is expensive. */
3049 	/* Make sure this is really our interrupt. */
3050 	if (!(CSR_READ_4(sc, BGE_MISC_LOCAL_CTL) & BGE_MLC_INTR_STATE))
3051 		return (0);
3052 #endif
3053 	/* Ack interrupt and stop others from occuring. */
3054 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 1);
3055 
3056 	BGE_EVCNT_INCR(sc->bge_ev_intr);
3057 
3058 	/*
3059 	 * Process link state changes.
3060 	 * Grrr. The link status word in the status block does
3061 	 * not work correctly on the BCM5700 rev AX and BX chips,
3062 	 * according to all available information. Hence, we have
3063 	 * to enable MII interrupts in order to properly obtain
3064 	 * async link changes. Unfortunately, this also means that
3065 	 * we have to read the MAC status register to detect link
3066 	 * changes, thereby adding an additional register access to
3067 	 * the interrupt handler.
3068 	 */
3069 
3070 	if (sc->bge_quirks & BGE_QUIRK_LINK_STATE_BROKEN) {
3071 		u_int32_t		status;
3072 
3073 		status = CSR_READ_4(sc, BGE_MAC_STS);
3074 		if (status & BGE_MACSTAT_MI_INTERRUPT) {
3075 			sc->bge_link = 0;
3076 			callout_stop(&sc->bge_timeout);
3077 			bge_tick(sc);
3078 			/* Clear the interrupt */
3079 			CSR_WRITE_4(sc, BGE_MAC_EVT_ENB,
3080 			    BGE_EVTENB_MI_INTERRUPT);
3081 			bge_miibus_readreg(&sc->bge_dev, 1, BRGPHY_MII_ISR);
3082 			bge_miibus_writereg(&sc->bge_dev, 1, BRGPHY_MII_IMR,
3083 			    BRGPHY_INTRS);
3084 		}
3085 	} else {
3086 		if (sc->bge_rdata->bge_status_block.bge_status &
3087 		    BGE_STATFLAG_LINKSTATE_CHANGED) {
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_STS, BGE_MACSTAT_SYNC_CHANGED|
3093 			    BGE_MACSTAT_CFG_CHANGED|BGE_MACSTAT_MI_COMPLETE|
3094 			    BGE_MACSTAT_LINK_CHANGED);
3095 		}
3096 	}
3097 
3098 	if (ifp->if_flags & IFF_RUNNING) {
3099 		/* Check RX return ring producer/consumer */
3100 		bge_rxeof(sc);
3101 
3102 		/* Check TX ring producer/consumer */
3103 		bge_txeof(sc);
3104 	}
3105 
3106 	if (sc->bge_pending_rxintr_change) {
3107 		uint32_t rx_ticks = sc->bge_rx_coal_ticks;
3108 		uint32_t rx_bds = sc->bge_rx_max_coal_bds;
3109 		uint32_t junk;
3110 
3111 		CSR_WRITE_4(sc, BGE_HCC_RX_COAL_TICKS, rx_ticks);
3112 		DELAY(10);
3113 		junk = CSR_READ_4(sc, BGE_HCC_RX_COAL_TICKS);
3114 
3115 		CSR_WRITE_4(sc, BGE_HCC_RX_MAX_COAL_BDS, rx_bds);
3116 		DELAY(10);
3117 		junk = CSR_READ_4(sc, BGE_HCC_RX_MAX_COAL_BDS);
3118 
3119 		sc->bge_pending_rxintr_change = 0;
3120 	}
3121 	bge_handle_events(sc);
3122 
3123 	/* Re-enable interrupts. */
3124 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 0);
3125 
3126 	if (ifp->if_flags & IFF_RUNNING && !IFQ_IS_EMPTY(&ifp->if_snd))
3127 		bge_start(ifp);
3128 
3129 	return (1);
3130 }
3131 
3132 static void
3133 bge_tick(void *xsc)
3134 {
3135 	struct bge_softc *sc = xsc;
3136 	struct mii_data *mii = &sc->bge_mii;
3137 	struct ifmedia *ifm = NULL;
3138 	struct ifnet *ifp = &sc->ethercom.ec_if;
3139 	int s;
3140 
3141 	s = splnet();
3142 
3143 	bge_stats_update(sc);
3144 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
3145 	if (sc->bge_link) {
3146 		splx(s);
3147 		return;
3148 	}
3149 
3150 	if (sc->bge_tbi) {
3151 		ifm = &sc->bge_ifmedia;
3152 		if (CSR_READ_4(sc, BGE_MAC_STS) &
3153 		    BGE_MACSTAT_TBI_PCS_SYNCHED) {
3154 			sc->bge_link++;
3155 			CSR_WRITE_4(sc, BGE_MAC_STS, 0xFFFFFFFF);
3156 			if (!IFQ_IS_EMPTY(&ifp->if_snd))
3157 				bge_start(ifp);
3158 		}
3159 		splx(s);
3160 		return;
3161 	}
3162 
3163 	mii_tick(mii);
3164 
3165 	if (!sc->bge_link && mii->mii_media_status & IFM_ACTIVE &&
3166 	    IFM_SUBTYPE(mii->mii_media_active) != IFM_NONE) {
3167 		sc->bge_link++;
3168 		if (!IFQ_IS_EMPTY(&ifp->if_snd))
3169 			bge_start(ifp);
3170 	}
3171 
3172 	splx(s);
3173 }
3174 
3175 static void
3176 bge_stats_update(struct bge_softc *sc)
3177 {
3178 	struct ifnet *ifp = &sc->ethercom.ec_if;
3179 	bus_size_t stats = BGE_MEMWIN_START + BGE_STATS_BLOCK;
3180 	bus_size_t rstats = BGE_RX_STATS;
3181 
3182 #define READ_RSTAT(sc, stats, stat) \
3183 	  CSR_READ_4(sc, stats + offsetof(struct bge_mac_stats_regs, stat))
3184 
3185 	if (sc->bge_quirks & BGE_QUIRK_5705_CORE) {
3186 		ifp->if_collisions +=
3187 		    READ_RSTAT(sc, rstats, dot3StatsSingleCollisionFrames) +
3188 		    READ_RSTAT(sc, rstats, dot3StatsMultipleCollisionFrames) +
3189 		    READ_RSTAT(sc, rstats, dot3StatsExcessiveCollisions) +
3190 		    READ_RSTAT(sc, rstats, dot3StatsLateCollisions);
3191 
3192 		BGE_EVCNT_ADD(sc->bge_ev_tx_xoff,
3193 			      READ_RSTAT(sc, rstats, outXoffSent));
3194 		BGE_EVCNT_ADD(sc->bge_ev_tx_xon,
3195 			      READ_RSTAT(sc, rstats, outXonSent));
3196 		BGE_EVCNT_ADD(sc->bge_ev_rx_xoff,
3197 			      READ_RSTAT(sc, rstats, xoffPauseFramesReceived));
3198 		BGE_EVCNT_ADD(sc->bge_ev_rx_xon,
3199 			      READ_RSTAT(sc, rstats, xonPauseFramesReceived));
3200 		BGE_EVCNT_ADD(sc->bge_ev_rx_macctl,
3201 			      READ_RSTAT(sc, rstats, macControlFramesReceived));
3202 		BGE_EVCNT_ADD(sc->bge_ev_xoffentered,
3203 			      READ_RSTAT(sc, rstats, xoffStateEntered));
3204 		return;
3205 	}
3206 
3207 #undef READ_RSTAT
3208 #define READ_STAT(sc, stats, stat) \
3209 	  CSR_READ_4(sc, stats + offsetof(struct bge_stats, stat))
3210 
3211 	ifp->if_collisions +=
3212 	  (READ_STAT(sc, stats, dot3StatsSingleCollisionFrames.bge_addr_lo) +
3213 	   READ_STAT(sc, stats, dot3StatsMultipleCollisionFrames.bge_addr_lo) +
3214 	   READ_STAT(sc, stats, dot3StatsExcessiveCollisions.bge_addr_lo) +
3215 	   READ_STAT(sc, stats, dot3StatsLateCollisions.bge_addr_lo)) -
3216 	  ifp->if_collisions;
3217 
3218 	BGE_EVCNT_UPD(sc->bge_ev_tx_xoff,
3219 		      READ_STAT(sc, stats, outXoffSent.bge_addr_lo));
3220 	BGE_EVCNT_UPD(sc->bge_ev_tx_xon,
3221 		      READ_STAT(sc, stats, outXonSent.bge_addr_lo));
3222 	BGE_EVCNT_UPD(sc->bge_ev_rx_xoff,
3223 		      READ_STAT(sc, stats,
3224 		      		xoffPauseFramesReceived.bge_addr_lo));
3225 	BGE_EVCNT_UPD(sc->bge_ev_rx_xon,
3226 		      READ_STAT(sc, stats, xonPauseFramesReceived.bge_addr_lo));
3227 	BGE_EVCNT_UPD(sc->bge_ev_rx_macctl,
3228 		      READ_STAT(sc, stats,
3229 		      		macControlFramesReceived.bge_addr_lo));
3230 	BGE_EVCNT_UPD(sc->bge_ev_xoffentered,
3231 		      READ_STAT(sc, stats, xoffStateEntered.bge_addr_lo));
3232 
3233 #undef READ_STAT
3234 
3235 #ifdef notdef
3236 	ifp->if_collisions +=
3237 	   (sc->bge_rdata->bge_info.bge_stats.dot3StatsSingleCollisionFrames +
3238 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsMultipleCollisionFrames +
3239 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsExcessiveCollisions +
3240 	   sc->bge_rdata->bge_info.bge_stats.dot3StatsLateCollisions) -
3241 	   ifp->if_collisions;
3242 #endif
3243 }
3244 
3245 /*
3246  * Pad outbound frame to ETHER_MIN_NOPAD for an unusual reason.
3247  * The bge hardware will pad out Tx runts to ETHER_MIN_NOPAD,
3248  * but when such padded frames employ the  bge IP/TCP checksum offload,
3249  * the hardware checksum assist gives incorrect results (possibly
3250  * from incorporating its own padding into the UDP/TCP checksum; who knows).
3251  * If we pad such runts with zeros, the onboard checksum comes out correct.
3252  */
3253 static inline int
3254 bge_cksum_pad(struct mbuf *pkt)
3255 {
3256 	struct mbuf *last = NULL;
3257 	int padlen;
3258 
3259 	padlen = ETHER_MIN_NOPAD - pkt->m_pkthdr.len;
3260 
3261 	/* if there's only the packet-header and we can pad there, use it. */
3262 	if (pkt->m_pkthdr.len == pkt->m_len &&
3263 	    !M_READONLY(pkt) && M_TRAILINGSPACE(pkt) >= padlen) {
3264 		last = pkt;
3265 	} else {
3266 		/*
3267 		 * Walk packet chain to find last mbuf. We will either
3268 		 * pad there, or append a new mbuf and pad it
3269 		 * (thus perhaps avoiding the bcm5700 dma-min bug).
3270 		 */
3271 		for (last = pkt; last->m_next != NULL; last = last->m_next) {
3272 	      	       (void) 0; /* do nothing*/
3273 		}
3274 
3275 		/* `last' now points to last in chain. */
3276 		if (!M_READONLY(last) && M_TRAILINGSPACE(last) >= padlen) {
3277 			(void) 0; /* we can pad here, in-place. */
3278 		} else {
3279 			/* Allocate new empty mbuf, pad it. Compact later. */
3280 			struct mbuf *n;
3281 			MGET(n, M_DONTWAIT, MT_DATA);
3282 			n->m_len = 0;
3283 			last->m_next = n;
3284 			last = n;
3285 		}
3286 	}
3287 
3288 #ifdef DEBUG
3289 	  /*KASSERT(M_WRITABLE(last), ("to-pad mbuf not writeable\n"));*/
3290 	  KASSERT(M_TRAILINGSPACE(last) >= padlen /*, ("insufficient space to pad\n")*/ );
3291 #endif
3292 	/* Now zero the pad area, to avoid the bge cksum-assist bug */
3293 	memset(mtod(last, caddr_t) + last->m_len, 0, padlen);
3294 	last->m_len += padlen;
3295 	pkt->m_pkthdr.len += padlen;
3296 	return 0;
3297 }
3298 
3299 /*
3300  * Compact outbound packets to avoid bug with DMA segments less than 8 bytes.
3301  */
3302 static inline int
3303 bge_compact_dma_runt(struct mbuf *pkt)
3304 {
3305 	struct mbuf	*m, *prev;
3306 	int 		totlen, prevlen;
3307 
3308 	prev = NULL;
3309 	totlen = 0;
3310 	prevlen = -1;
3311 
3312 	for (m = pkt; m != NULL; prev = m,m = m->m_next) {
3313 		int mlen = m->m_len;
3314 		int shortfall = 8 - mlen ;
3315 
3316 		totlen += mlen;
3317 		if (mlen == 0) {
3318 			continue;
3319 		}
3320 		if (mlen >= 8)
3321 			continue;
3322 
3323 		/* If we get here, mbuf data is too small for DMA engine.
3324 		 * Try to fix by shuffling data to prev or next in chain.
3325 		 * If that fails, do a compacting deep-copy of the whole chain.
3326 		 */
3327 
3328 		/* Internal frag. If fits in prev, copy it there. */
3329 		if (prev && !M_READONLY(prev) &&
3330 		      M_TRAILINGSPACE(prev) >= m->m_len) {
3331 		  	bcopy(m->m_data,
3332 			      prev->m_data+prev->m_len,
3333 			      mlen);
3334 			prev->m_len += mlen;
3335 			m->m_len = 0;
3336 			/* XXX stitch chain */
3337 			prev->m_next = m_free(m);
3338 			m = prev;
3339 			continue;
3340 		}
3341 		else if (m->m_next != NULL && !M_READONLY(m) &&
3342 			     M_TRAILINGSPACE(m) >= shortfall &&
3343 			     m->m_next->m_len >= (8 + shortfall)) {
3344 		    /* m is writable and have enough data in next, pull up. */
3345 
3346 		  	bcopy(m->m_next->m_data,
3347 			      m->m_data+m->m_len,
3348 			      shortfall);
3349 			m->m_len += shortfall;
3350 			m->m_next->m_len -= shortfall;
3351 			m->m_next->m_data += shortfall;
3352 		}
3353 		else if (m->m_next == NULL || 1) {
3354 		  	/* Got a runt at the very end of the packet.
3355 			 * borrow data from the tail of the preceding mbuf and
3356 			 * update its length in-place. (The original data is still
3357 			 * valid, so we can do this even if prev is not writable.)
3358 			 */
3359 
3360 			/* if we'd make prev a runt, just move all of its data. */
3361 #ifdef DEBUG
3362 			KASSERT(prev != NULL /*, ("runt but null PREV")*/);
3363 			KASSERT(prev->m_len >= 8 /*, ("runt prev")*/);
3364 #endif
3365 			if ((prev->m_len - shortfall) < 8)
3366 				shortfall = prev->m_len;
3367 
3368 #ifdef notyet	/* just do the safe slow thing for now */
3369 			if (!M_READONLY(m)) {
3370 				if (M_LEADINGSPACE(m) < shorfall) {
3371 					void *m_dat;
3372 					m_dat = (m->m_flags & M_PKTHDR) ?
3373 					  m->m_pktdat : m->dat;
3374 					memmove(m_dat, mtod(m, void*), m->m_len);
3375 					m->m_data = m_dat;
3376 				    }
3377 			} else
3378 #endif	/* just do the safe slow thing */
3379 			{
3380 				struct mbuf * n = NULL;
3381 				int newprevlen = prev->m_len - shortfall;
3382 
3383 				MGET(n, M_NOWAIT, MT_DATA);
3384 				if (n == NULL)
3385 				   return ENOBUFS;
3386 				KASSERT(m->m_len + shortfall < MLEN
3387 					/*,
3388 					  ("runt %d +prev %d too big\n", m->m_len, shortfall)*/);
3389 
3390 				/* first copy the data we're stealing from prev */
3391 				bcopy(prev->m_data + newprevlen, n->m_data, shortfall);
3392 
3393 				/* update prev->m_len accordingly */
3394 				prev->m_len -= shortfall;
3395 
3396 				/* copy data from runt m */
3397 				bcopy(m->m_data, n->m_data + shortfall, m->m_len);
3398 
3399 				/* n holds what we stole from prev, plus m */
3400 				n->m_len = shortfall + m->m_len;
3401 
3402 				/* stitch n into chain and free m */
3403 				n->m_next = m->m_next;
3404 				prev->m_next = n;
3405 				/* KASSERT(m->m_next == NULL); */
3406 				m->m_next = NULL;
3407 				m_free(m);
3408 				m = n;	/* for continuing loop */
3409 			}
3410 		}
3411 		prevlen = m->m_len;
3412 	}
3413 	return 0;
3414 }
3415 
3416 /*
3417  * Encapsulate an mbuf chain in the tx ring  by coupling the mbuf data
3418  * pointers to descriptors.
3419  */
3420 static int
3421 bge_encap(struct bge_softc *sc, struct mbuf *m_head, u_int32_t *txidx)
3422 {
3423 	struct bge_tx_bd	*f = NULL;
3424 	u_int32_t		frag, cur, cnt = 0;
3425 	u_int16_t		csum_flags = 0;
3426 	u_int16_t		txbd_tso_flags = 0;
3427 	struct txdmamap_pool_entry *dma;
3428 	bus_dmamap_t dmamap;
3429 	int			i = 0;
3430 	struct m_tag		*mtag;
3431 	int			use_tso, maxsegsize, error;
3432 
3433 	cur = frag = *txidx;
3434 
3435 	if (m_head->m_pkthdr.csum_flags) {
3436 		if (m_head->m_pkthdr.csum_flags & M_CSUM_IPv4)
3437 			csum_flags |= BGE_TXBDFLAG_IP_CSUM;
3438 		if (m_head->m_pkthdr.csum_flags & (M_CSUM_TCPv4|M_CSUM_UDPv4))
3439 			csum_flags |= BGE_TXBDFLAG_TCP_UDP_CSUM;
3440 	}
3441 
3442 	/*
3443 	 * If we were asked to do an outboard checksum, and the NIC
3444 	 * has the bug where it sometimes adds in the Ethernet padding,
3445 	 * explicitly pad with zeros so the cksum will be correct either way.
3446 	 * (For now, do this for all chip versions, until newer
3447 	 * are confirmed to not require the workaround.)
3448 	 */
3449 	if ((csum_flags & BGE_TXBDFLAG_TCP_UDP_CSUM) == 0 ||
3450 #ifdef notyet
3451 	    (sc->bge_quirks & BGE_QUIRK_SHORT_CKSUM_BUG) == 0 ||
3452 #endif
3453 	    m_head->m_pkthdr.len >= ETHER_MIN_NOPAD)
3454 		goto check_dma_bug;
3455 
3456 	if (bge_cksum_pad(m_head) != 0) {
3457 	    return ENOBUFS;
3458 	}
3459 
3460 check_dma_bug:
3461 	if (!(sc->bge_quirks & BGE_QUIRK_5700_SMALLDMA))
3462 		goto doit;
3463 	/*
3464 	 * bcm5700 Revision B silicon cannot handle DMA descriptors with
3465 	 * less than eight bytes.  If we encounter a teeny mbuf
3466 	 * at the end of a chain, we can pad.  Otherwise, copy.
3467 	 */
3468 	if (bge_compact_dma_runt(m_head) != 0)
3469 		return ENOBUFS;
3470 
3471 doit:
3472 	dma = SLIST_FIRST(&sc->txdma_list);
3473 	if (dma == NULL)
3474 		return ENOBUFS;
3475 	dmamap = dma->dmamap;
3476 
3477 	/*
3478 	 * Set up any necessary TSO state before we start packing...
3479 	 */
3480 	use_tso = (m_head->m_pkthdr.csum_flags & M_CSUM_TSOv4) != 0;
3481 	if (!use_tso) {
3482 		maxsegsize = 0;
3483 	} else {	/* TSO setup */
3484 		unsigned  mss;
3485 		struct ether_header *eh;
3486 		unsigned ip_tcp_hlen, iptcp_opt_words, tcp_seg_flags, offset;
3487 		struct mbuf * m0 = m_head;
3488 		struct ip *ip;
3489 		struct tcphdr *th;
3490 		int iphl, hlen;
3491 
3492 		/*
3493 		 * XXX It would be nice if the mbuf pkthdr had offset
3494 		 * fields for the protocol headers.
3495 		 */
3496 
3497 		eh = mtod(m0, struct ether_header *);
3498 		switch (htons(eh->ether_type)) {
3499 		case ETHERTYPE_IP:
3500 			offset = ETHER_HDR_LEN;
3501 			break;
3502 
3503 		case ETHERTYPE_VLAN:
3504 			offset = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3505 			break;
3506 
3507 		default:
3508 			/*
3509 			 * Don't support this protocol or encapsulation.
3510 			 */
3511 			return (ENOBUFS);
3512 		}
3513 
3514 		/*
3515 		 * TCP/IP headers are in the first mbuf; we can do
3516 		 * this the easy way.
3517 		 */
3518 		iphl = M_CSUM_DATA_IPv4_IPHL(m0->m_pkthdr.csum_data);
3519 		hlen = iphl + offset;
3520 		if (__predict_false(m0->m_len <
3521 				    (hlen + sizeof(struct tcphdr)))) {
3522 
3523 			  printf("TSO: hard case m0->m_len == %d <"
3524 				 " ip/tcp hlen %zd, not handled yet\n",
3525 				 m0->m_len, hlen+ sizeof(struct tcphdr));
3526 #ifdef NOTYET
3527 			/*
3528 			 * XXX jonathan@NetBSD.org: untested.
3529 			 * how to force  this branch to be taken?
3530 			 */
3531 			BGE_EVCNT_INCR(&sc->sc_ev_txtsopain);
3532 
3533 			m_copydata(m0, offset, sizeof(ip), &ip);
3534 			m_copydata(m0, hlen, sizeof(th), &th);
3535 
3536 			ip.ip_len = 0;
3537 
3538 			m_copyback(m0, hlen + offsetof(struct ip, ip_len),
3539 			    sizeof(ip.ip_len), &ip.ip_len);
3540 
3541 			th.th_sum = in_cksum_phdr(ip.ip_src.s_addr,
3542 			    ip.ip_dst.s_addr, htons(IPPROTO_TCP));
3543 
3544 			m_copyback(m0, hlen + offsetof(struct tcphdr, th_sum),
3545 			    sizeof(th.th_sum), &th.th_sum);
3546 
3547 			hlen += th.th_off << 2;
3548 			iptcp_opt_words	= hlen;
3549 #else
3550 			/*
3551 			 * if_wm "hard" case not yet supported, can we not
3552 			 * mandate it out of existence?
3553 			 */
3554 			(void) ip; (void)th; (void) ip_tcp_hlen;
3555 
3556 			return ENOBUFS;
3557 #endif
3558 		} else {
3559 			ip = (struct ip *) (mtod(m0, caddr_t) + offset);
3560 			th = (struct tcphdr *) (mtod(m0, caddr_t) + hlen);
3561 			ip_tcp_hlen = iphl +  (th->th_off << 2);
3562 
3563 			/* Total IP/TCP options, in 32-bit words */
3564 			iptcp_opt_words = (ip_tcp_hlen
3565 					   - sizeof(struct tcphdr)
3566 					   - sizeof(struct ip)) >> 2;
3567 		}
3568 		if (BGE_IS_5750_OR_BEYOND(sc)) {
3569 			th->th_sum = 0;
3570 			csum_flags &= ~(BGE_TXBDFLAG_TCP_UDP_CSUM);
3571 		} else {
3572 			/*
3573 			 * XXX jonathan@NetBSD.org: 5705 untested.
3574 			 * Requires TSO firmware patch for 5701/5703/5704.
3575 			 */
3576 			th->th_sum = in_cksum_phdr(ip->ip_src.s_addr,
3577 			    ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3578 		}
3579 
3580 		mss = m_head->m_pkthdr.segsz;
3581 		txbd_tso_flags |=
3582 		    BGE_TXBDFLAG_CPU_PRE_DMA |
3583 		    BGE_TXBDFLAG_CPU_POST_DMA;
3584 
3585 		/*
3586 		 * Our NIC TSO-assist assumes TSO has standard, optionless
3587 		 * IPv4 and TCP headers, which total 40 bytes. By default,
3588 		 * the NIC copies 40 bytes of IP/TCP header from the
3589 		 * supplied header into the IP/TCP header portion of
3590 		 * each post-TSO-segment. If the supplied packet has IP or
3591 		 * TCP options, we need to tell the NIC to copy those extra
3592 		 * bytes into each  post-TSO header, in addition to the normal
3593 		 * 40-byte IP/TCP header (and to leave space accordingly).
3594 		 * Unfortunately, the driver encoding of option length
3595 		 * varies across different ASIC families.
3596 		 */
3597 		tcp_seg_flags = 0;
3598 		if (iptcp_opt_words) {
3599 			if ( BGE_IS_5705_OR_BEYOND(sc)) {
3600 				tcp_seg_flags =
3601 					iptcp_opt_words << 11;
3602 			} else {
3603 				txbd_tso_flags |=
3604 					iptcp_opt_words << 12;
3605 			}
3606 		}
3607 		maxsegsize = mss | tcp_seg_flags;
3608 		ip->ip_len = htons(mss + ip_tcp_hlen);
3609 
3610 	}	/* TSO setup */
3611 
3612 	/*
3613 	 * Start packing the mbufs in this chain into
3614 	 * the fragment pointers. Stop when we run out
3615 	 * of fragments or hit the end of the mbuf chain.
3616 	 */
3617 	error = bus_dmamap_load_mbuf(sc->bge_dmatag, dmamap, m_head,
3618 	    BUS_DMA_NOWAIT);
3619 	if (error) {
3620 		return(ENOBUFS);
3621 	}
3622 
3623 	mtag = sc->ethercom.ec_nvlans ?
3624 	    m_tag_find(m_head, PACKET_TAG_VLAN, NULL) : NULL;
3625 
3626 
3627 	/* Iterate over dmap-map fragments. */
3628 	for (i = 0; i < dmamap->dm_nsegs; i++) {
3629 		f = &sc->bge_rdata->bge_tx_ring[frag];
3630 		if (sc->bge_cdata.bge_tx_chain[frag] != NULL)
3631 			break;
3632 
3633 		bge_set_hostaddr(&f->bge_addr, dmamap->dm_segs[i].ds_addr);
3634 		f->bge_len = dmamap->dm_segs[i].ds_len;
3635 
3636 		/*
3637 		 * For 5751 and follow-ons, for TSO we must turn
3638 		 * off checksum-assist flag in the tx-descr, and
3639 		 * supply the ASIC-revision-specific encoding
3640 		 * of TSO flags and segsize.
3641 		 */
3642 		if (use_tso) {
3643 			if (BGE_IS_5750_OR_BEYOND(sc) || i == 0) {
3644 				f->bge_rsvd = maxsegsize;
3645 				f->bge_flags = csum_flags | txbd_tso_flags;
3646 			} else {
3647 				f->bge_rsvd = 0;
3648 				f->bge_flags =
3649 				  (csum_flags | txbd_tso_flags) & 0x0fff;
3650 			}
3651 		} else {
3652 			f->bge_rsvd = 0;
3653 			f->bge_flags = csum_flags;
3654 		}
3655 
3656 		if (mtag != NULL) {
3657 			f->bge_flags |= BGE_TXBDFLAG_VLAN_TAG;
3658 			f->bge_vlan_tag = VLAN_TAG_VALUE(mtag);
3659 		} else {
3660 			f->bge_vlan_tag = 0;
3661 		}
3662 		/*
3663 		 * Sanity check: avoid coming within 16 descriptors
3664 		 * of the end of the ring.
3665 		 */
3666 		if ((BGE_TX_RING_CNT - (sc->bge_txcnt + cnt)) < 16) {
3667 			BGE_TSO_PRINTF(("%s: "
3668 			    " dmamap_load_mbuf too close to ring wrap\n",
3669 			    sc->bge_dev.dv_xname));
3670 			return(ENOBUFS);
3671 		}
3672 		cur = frag;
3673 		BGE_INC(frag, BGE_TX_RING_CNT);
3674 		cnt++;
3675 	}
3676 
3677 	if (i < dmamap->dm_nsegs) {
3678 		BGE_TSO_PRINTF(("%s: reached %d < dm_nsegs %d\n",
3679 		    sc->bge_dev.dv_xname, i, dmamap->dm_nsegs));
3680 		return ENOBUFS;
3681 	}
3682 
3683 	bus_dmamap_sync(sc->bge_dmatag, dmamap, 0, dmamap->dm_mapsize,
3684 	    BUS_DMASYNC_PREWRITE);
3685 
3686 	if (frag == sc->bge_tx_saved_considx) {
3687 		BGE_TSO_PRINTF(("%s: frag %d = wrapped id %d?\n",
3688 		    sc->bge_dev.dv_xname, frag, sc->bge_tx_saved_considx));
3689 
3690 		return(ENOBUFS);
3691 	}
3692 
3693 	sc->bge_rdata->bge_tx_ring[cur].bge_flags |= BGE_TXBDFLAG_END;
3694 	sc->bge_cdata.bge_tx_chain[cur] = m_head;
3695 	SLIST_REMOVE_HEAD(&sc->txdma_list, link);
3696 	sc->txdma[cur] = dma;
3697 	sc->bge_txcnt += cnt;
3698 
3699 	*txidx = frag;
3700 
3701 	return(0);
3702 }
3703 
3704 /*
3705  * Main transmit routine. To avoid having to do mbuf copies, we put pointers
3706  * to the mbuf data regions directly in the transmit descriptors.
3707  */
3708 static void
3709 bge_start(struct ifnet *ifp)
3710 {
3711 	struct bge_softc *sc;
3712 	struct mbuf *m_head = NULL;
3713 	u_int32_t prodidx;
3714 	int pkts = 0;
3715 
3716 	sc = ifp->if_softc;
3717 
3718 	if (!sc->bge_link && ifp->if_snd.ifq_len < 10)
3719 		return;
3720 
3721 	prodidx = sc->bge_tx_prodidx;
3722 
3723 	while(sc->bge_cdata.bge_tx_chain[prodidx] == NULL) {
3724 		IFQ_POLL(&ifp->if_snd, m_head);
3725 		if (m_head == NULL)
3726 			break;
3727 
3728 #if 0
3729 		/*
3730 		 * XXX
3731 		 * safety overkill.  If this is a fragmented packet chain
3732 		 * with delayed TCP/UDP checksums, then only encapsulate
3733 		 * it if we have enough descriptors to handle the entire
3734 		 * chain at once.
3735 		 * (paranoia -- may not actually be needed)
3736 		 */
3737 		if (m_head->m_flags & M_FIRSTFRAG &&
3738 		    m_head->m_pkthdr.csum_flags & (CSUM_DELAY_DATA)) {
3739 			if ((BGE_TX_RING_CNT - sc->bge_txcnt) <
3740 			    M_CSUM_DATA_IPv4_OFFSET(m_head->m_pkthdr.csum_data) + 16) {
3741 				ifp->if_flags |= IFF_OACTIVE;
3742 				break;
3743 			}
3744 		}
3745 #endif
3746 
3747 		/*
3748 		 * Pack the data into the transmit ring. If we
3749 		 * don't have room, set the OACTIVE flag and wait
3750 		 * for the NIC to drain the ring.
3751 		 */
3752 		if (bge_encap(sc, m_head, &prodidx)) {
3753 			printf("bge: failed on len %d?\n", m_head->m_pkthdr.len);
3754 			ifp->if_flags |= IFF_OACTIVE;
3755 			break;
3756 		}
3757 
3758 		/* now we are committed to transmit the packet */
3759 		IFQ_DEQUEUE(&ifp->if_snd, m_head);
3760 		pkts++;
3761 
3762 #if NBPFILTER > 0
3763 		/*
3764 		 * If there's a BPF listener, bounce a copy of this frame
3765 		 * to him.
3766 		 */
3767 		if (ifp->if_bpf)
3768 			bpf_mtap(ifp->if_bpf, m_head);
3769 #endif
3770 	}
3771 	if (pkts == 0)
3772 		return;
3773 
3774 	/* Transmit */
3775 	CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
3776 	if (sc->bge_quirks & BGE_QUIRK_PRODUCER_BUG)	/* 5700 b2 errata */
3777 		CSR_WRITE_4(sc, BGE_MBX_TX_HOST_PROD0_LO, prodidx);
3778 
3779 	sc->bge_tx_prodidx = prodidx;
3780 
3781 	/*
3782 	 * Set a timeout in case the chip goes out to lunch.
3783 	 */
3784 	ifp->if_timer = 5;
3785 }
3786 
3787 static int
3788 bge_init(struct ifnet *ifp)
3789 {
3790 	struct bge_softc *sc = ifp->if_softc;
3791 	u_int16_t *m;
3792 	int s, error;
3793 
3794 	s = splnet();
3795 
3796 	ifp = &sc->ethercom.ec_if;
3797 
3798 	/* Cancel pending I/O and flush buffers. */
3799 	bge_stop(sc);
3800 	bge_reset(sc);
3801 	bge_chipinit(sc);
3802 
3803 	/*
3804 	 * Init the various state machines, ring
3805 	 * control blocks and firmware.
3806 	 */
3807 	error = bge_blockinit(sc);
3808 	if (error != 0) {
3809 		printf("%s: initialization error %d\n", sc->bge_dev.dv_xname,
3810 		    error);
3811 		splx(s);
3812 		return error;
3813 	}
3814 
3815 	ifp = &sc->ethercom.ec_if;
3816 
3817 	/* Specify MTU. */
3818 	CSR_WRITE_4(sc, BGE_RX_MTU, ifp->if_mtu +
3819 	    ETHER_HDR_LEN + ETHER_CRC_LEN + ETHER_VLAN_ENCAP_LEN);
3820 
3821 	/* Load our MAC address. */
3822 	m = (u_int16_t *)&(LLADDR(ifp->if_sadl)[0]);
3823 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_LO, htons(m[0]));
3824 	CSR_WRITE_4(sc, BGE_MAC_ADDR1_HI, (htons(m[1]) << 16) | htons(m[2]));
3825 
3826 	/* Enable or disable promiscuous mode as needed. */
3827 	if (ifp->if_flags & IFF_PROMISC) {
3828 		BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
3829 	} else {
3830 		BGE_CLRBIT(sc, BGE_RX_MODE, BGE_RXMODE_RX_PROMISC);
3831 	}
3832 
3833 	/* Program multicast filter. */
3834 	bge_setmulti(sc);
3835 
3836 	/* Init RX ring. */
3837 	bge_init_rx_ring_std(sc);
3838 
3839 	/* Init jumbo RX ring. */
3840 	if (ifp->if_mtu > (ETHERMTU + ETHER_HDR_LEN + ETHER_CRC_LEN))
3841 		bge_init_rx_ring_jumbo(sc);
3842 
3843 	/* Init our RX return ring index */
3844 	sc->bge_rx_saved_considx = 0;
3845 
3846 	/* Init TX ring. */
3847 	bge_init_tx_ring(sc);
3848 
3849 	/* Turn on transmitter */
3850 	BGE_SETBIT(sc, BGE_TX_MODE, BGE_TXMODE_ENABLE);
3851 
3852 	/* Turn on receiver */
3853 	BGE_SETBIT(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
3854 
3855 	CSR_WRITE_4(sc, BGE_MAX_RX_FRAME_LOWAT, 2);
3856 
3857 	/* Tell firmware we're alive. */
3858 	BGE_SETBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
3859 
3860 	/* Enable host interrupts. */
3861 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_CLEAR_INTA);
3862 	BGE_CLRBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
3863 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 0);
3864 
3865 	bge_ifmedia_upd(ifp);
3866 
3867 	ifp->if_flags |= IFF_RUNNING;
3868 	ifp->if_flags &= ~IFF_OACTIVE;
3869 
3870 	splx(s);
3871 
3872 	callout_reset(&sc->bge_timeout, hz, bge_tick, sc);
3873 
3874 	return 0;
3875 }
3876 
3877 /*
3878  * Set media options.
3879  */
3880 static int
3881 bge_ifmedia_upd(struct ifnet *ifp)
3882 {
3883 	struct bge_softc *sc = ifp->if_softc;
3884 	struct mii_data *mii = &sc->bge_mii;
3885 	struct ifmedia *ifm = &sc->bge_ifmedia;
3886 
3887 	/* If this is a 1000baseX NIC, enable the TBI port. */
3888 	if (sc->bge_tbi) {
3889 		if (IFM_TYPE(ifm->ifm_media) != IFM_ETHER)
3890 			return(EINVAL);
3891 		switch(IFM_SUBTYPE(ifm->ifm_media)) {
3892 		case IFM_AUTO:
3893 			break;
3894 		case IFM_1000_SX:
3895 			if ((ifm->ifm_media & IFM_GMASK) == IFM_FDX) {
3896 				BGE_CLRBIT(sc, BGE_MAC_MODE,
3897 				    BGE_MACMODE_HALF_DUPLEX);
3898 			} else {
3899 				BGE_SETBIT(sc, BGE_MAC_MODE,
3900 				    BGE_MACMODE_HALF_DUPLEX);
3901 			}
3902 			break;
3903 		default:
3904 			return(EINVAL);
3905 		}
3906 		/* XXX 802.3x flow control for 1000BASE-SX */
3907 		return(0);
3908 	}
3909 
3910 	sc->bge_link = 0;
3911 	mii_mediachg(mii);
3912 
3913 	return(0);
3914 }
3915 
3916 /*
3917  * Report current media status.
3918  */
3919 static void
3920 bge_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
3921 {
3922 	struct bge_softc *sc = ifp->if_softc;
3923 	struct mii_data *mii = &sc->bge_mii;
3924 
3925 	if (sc->bge_tbi) {
3926 		ifmr->ifm_status = IFM_AVALID;
3927 		ifmr->ifm_active = IFM_ETHER;
3928 		if (CSR_READ_4(sc, BGE_MAC_STS) &
3929 		    BGE_MACSTAT_TBI_PCS_SYNCHED)
3930 			ifmr->ifm_status |= IFM_ACTIVE;
3931 		ifmr->ifm_active |= IFM_1000_SX;
3932 		if (CSR_READ_4(sc, BGE_MAC_MODE) & BGE_MACMODE_HALF_DUPLEX)
3933 			ifmr->ifm_active |= IFM_HDX;
3934 		else
3935 			ifmr->ifm_active |= IFM_FDX;
3936 		return;
3937 	}
3938 
3939 	mii_pollstat(mii);
3940 	ifmr->ifm_status = mii->mii_media_status;
3941 	ifmr->ifm_active = (mii->mii_media_active & ~IFM_ETH_FMASK) |
3942 	    sc->bge_flowflags;
3943 }
3944 
3945 static int
3946 bge_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
3947 {
3948 	struct bge_softc *sc = ifp->if_softc;
3949 	struct ifreq *ifr = (struct ifreq *) data;
3950 	int s, error = 0;
3951 	struct mii_data *mii;
3952 
3953 	s = splnet();
3954 
3955 	switch(command) {
3956 	case SIOCSIFFLAGS:
3957 		if (ifp->if_flags & IFF_UP) {
3958 			/*
3959 			 * If only the state of the PROMISC flag changed,
3960 			 * then just use the 'set promisc mode' command
3961 			 * instead of reinitializing the entire NIC. Doing
3962 			 * a full re-init means reloading the firmware and
3963 			 * waiting for it to start up, which may take a
3964 			 * second or two.
3965 			 */
3966 			if (ifp->if_flags & IFF_RUNNING &&
3967 			    ifp->if_flags & IFF_PROMISC &&
3968 			    !(sc->bge_if_flags & IFF_PROMISC)) {
3969 				BGE_SETBIT(sc, BGE_RX_MODE,
3970 				    BGE_RXMODE_RX_PROMISC);
3971 			} else if (ifp->if_flags & IFF_RUNNING &&
3972 			    !(ifp->if_flags & IFF_PROMISC) &&
3973 			    sc->bge_if_flags & IFF_PROMISC) {
3974 				BGE_CLRBIT(sc, BGE_RX_MODE,
3975 				    BGE_RXMODE_RX_PROMISC);
3976 			} else if (!(sc->bge_if_flags & IFF_UP))
3977 				bge_init(ifp);
3978 		} else {
3979 			if (ifp->if_flags & IFF_RUNNING) {
3980 				bge_stop(sc);
3981 			}
3982 		}
3983 		sc->bge_if_flags = ifp->if_flags;
3984 		error = 0;
3985 		break;
3986 	case SIOCSIFMEDIA:
3987 		/* XXX Flow control is not supported for 1000BASE-SX */
3988 		if (sc->bge_tbi) {
3989 			ifr->ifr_media &= ~IFM_ETH_FMASK;
3990 			sc->bge_flowflags = 0;
3991 		}
3992 
3993 		/* Flow control requires full-duplex mode. */
3994 		if (IFM_SUBTYPE(ifr->ifr_media) == IFM_AUTO ||
3995 		    (ifr->ifr_media & IFM_FDX) == 0) {
3996 		    	ifr->ifr_media &= ~IFM_ETH_FMASK;
3997 		}
3998 		if (IFM_SUBTYPE(ifr->ifr_media) != IFM_AUTO) {
3999 			if ((ifr->ifr_media & IFM_ETH_FMASK) == IFM_FLOW) {
4000 				/* We an do both TXPAUSE and RXPAUSE. */
4001 				ifr->ifr_media |=
4002 				    IFM_ETH_TXPAUSE | IFM_ETH_RXPAUSE;
4003 			}
4004 			sc->bge_flowflags = ifr->ifr_media & IFM_ETH_FMASK;
4005 		}
4006 		/* FALLTHROUGH */
4007 	case SIOCGIFMEDIA:
4008 		if (sc->bge_tbi) {
4009 			error = ifmedia_ioctl(ifp, ifr, &sc->bge_ifmedia,
4010 			    command);
4011 		} else {
4012 			mii = &sc->bge_mii;
4013 			error = ifmedia_ioctl(ifp, ifr, &mii->mii_media,
4014 			    command);
4015 		}
4016 		break;
4017 	default:
4018 		error = ether_ioctl(ifp, command, data);
4019 		if (error == ENETRESET) {
4020 			if (ifp->if_flags & IFF_RUNNING)
4021 				bge_setmulti(sc);
4022 			error = 0;
4023 		}
4024 		break;
4025 	}
4026 
4027 	splx(s);
4028 
4029 	return(error);
4030 }
4031 
4032 static void
4033 bge_watchdog(struct ifnet *ifp)
4034 {
4035 	struct bge_softc *sc;
4036 
4037 	sc = ifp->if_softc;
4038 
4039 	printf("%s: watchdog timeout -- resetting\n", sc->bge_dev.dv_xname);
4040 
4041 	ifp->if_flags &= ~IFF_RUNNING;
4042 	bge_init(ifp);
4043 
4044 	ifp->if_oerrors++;
4045 }
4046 
4047 static void
4048 bge_stop_block(struct bge_softc *sc, bus_addr_t reg, uint32_t bit)
4049 {
4050 	int i;
4051 
4052 	BGE_CLRBIT(sc, reg, bit);
4053 
4054 	for (i = 0; i < BGE_TIMEOUT; i++) {
4055 		if ((CSR_READ_4(sc, reg) & bit) == 0)
4056 			return;
4057 		delay(100);
4058 		if (sc->bge_pcie)
4059 		  DELAY(1000);
4060 	}
4061 
4062 	printf("%s: block failed to stop: reg 0x%lx, bit 0x%08x\n",
4063 	    sc->bge_dev.dv_xname, (u_long) reg, bit);
4064 }
4065 
4066 /*
4067  * Stop the adapter and free any mbufs allocated to the
4068  * RX and TX lists.
4069  */
4070 static void
4071 bge_stop(struct bge_softc *sc)
4072 {
4073 	struct ifnet *ifp = &sc->ethercom.ec_if;
4074 
4075 	callout_stop(&sc->bge_timeout);
4076 
4077 	/*
4078 	 * Disable all of the receiver blocks
4079 	 */
4080 	bge_stop_block(sc, BGE_RX_MODE, BGE_RXMODE_ENABLE);
4081 	bge_stop_block(sc, BGE_RBDI_MODE, BGE_RBDIMODE_ENABLE);
4082 	bge_stop_block(sc, BGE_RXLP_MODE, BGE_RXLPMODE_ENABLE);
4083 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4084 		bge_stop_block(sc, BGE_RXLS_MODE, BGE_RXLSMODE_ENABLE);
4085 	}
4086 	bge_stop_block(sc, BGE_RDBDI_MODE, BGE_RBDIMODE_ENABLE);
4087 	bge_stop_block(sc, BGE_RDC_MODE, BGE_RDCMODE_ENABLE);
4088 	bge_stop_block(sc, BGE_RBDC_MODE, BGE_RBDCMODE_ENABLE);
4089 
4090 	/*
4091 	 * Disable all of the transmit blocks
4092 	 */
4093 	bge_stop_block(sc, BGE_SRS_MODE, BGE_SRSMODE_ENABLE);
4094 	bge_stop_block(sc, BGE_SBDI_MODE, BGE_SBDIMODE_ENABLE);
4095 	bge_stop_block(sc, BGE_SDI_MODE, BGE_SDIMODE_ENABLE);
4096 	bge_stop_block(sc, BGE_RDMA_MODE, BGE_RDMAMODE_ENABLE);
4097 	bge_stop_block(sc, BGE_SDC_MODE, BGE_SDCMODE_ENABLE);
4098 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4099 		bge_stop_block(sc, BGE_DMAC_MODE, BGE_DMACMODE_ENABLE);
4100 	}
4101 	bge_stop_block(sc, BGE_SBDC_MODE, BGE_SBDCMODE_ENABLE);
4102 
4103 	/*
4104 	 * Shut down all of the memory managers and related
4105 	 * state machines.
4106 	 */
4107 	bge_stop_block(sc, BGE_HCC_MODE, BGE_HCCMODE_ENABLE);
4108 	bge_stop_block(sc, BGE_WDMA_MODE, BGE_WDMAMODE_ENABLE);
4109 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4110 		bge_stop_block(sc, BGE_MBCF_MODE, BGE_MBCFMODE_ENABLE);
4111 	}
4112 
4113 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0xFFFFFFFF);
4114 	CSR_WRITE_4(sc, BGE_FTQ_RESET, 0);
4115 
4116 	if ((sc->bge_quirks & BGE_QUIRK_5705_CORE) == 0) {
4117 		bge_stop_block(sc, BGE_BMAN_MODE, BGE_BMANMODE_ENABLE);
4118 		bge_stop_block(sc, BGE_MARB_MODE, BGE_MARBMODE_ENABLE);
4119 	}
4120 
4121 	/* Disable host interrupts. */
4122 	BGE_SETBIT(sc, BGE_PCI_MISC_CTL, BGE_PCIMISCCTL_MASK_PCI_INTR);
4123 	CSR_WRITE_4(sc, BGE_MBX_IRQ0_LO, 1);
4124 
4125 	/*
4126 	 * Tell firmware we're shutting down.
4127 	 */
4128 	BGE_CLRBIT(sc, BGE_MODE_CTL, BGE_MODECTL_STACKUP);
4129 
4130 	/* Free the RX lists. */
4131 	bge_free_rx_ring_std(sc);
4132 
4133 	/* Free jumbo RX list. */
4134 	bge_free_rx_ring_jumbo(sc);
4135 
4136 	/* Free TX buffers. */
4137 	bge_free_tx_ring(sc);
4138 
4139 	/*
4140 	 * Isolate/power down the PHY.
4141 	 */
4142 	if (!sc->bge_tbi)
4143 		mii_down(&sc->bge_mii);
4144 
4145 	sc->bge_link = 0;
4146 
4147 	sc->bge_tx_saved_considx = BGE_TXCONS_UNSET;
4148 
4149 	ifp->if_flags &= ~(IFF_RUNNING | IFF_OACTIVE);
4150 }
4151 
4152 /*
4153  * Stop all chip I/O so that the kernel's probe routines don't
4154  * get confused by errant DMAs when rebooting.
4155  */
4156 static void
4157 bge_shutdown(void *xsc)
4158 {
4159 	struct bge_softc *sc = (struct bge_softc *)xsc;
4160 
4161 	bge_stop(sc);
4162 	bge_reset(sc);
4163 }
4164 
4165 
4166 static int
4167 sysctl_bge_verify(SYSCTLFN_ARGS)
4168 {
4169 	int error, t;
4170 	struct sysctlnode node;
4171 
4172 	node = *rnode;
4173 	t = *(int*)rnode->sysctl_data;
4174 	node.sysctl_data = &t;
4175 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
4176 	if (error || newp == NULL)
4177 		return (error);
4178 
4179 #if 0
4180 	DPRINTF2(("%s: t = %d, nodenum = %d, rnodenum = %d\n", __func__, t,
4181 	    node.sysctl_num, rnode->sysctl_num));
4182 #endif
4183 
4184 	if (node.sysctl_num == bge_rxthresh_nodenum) {
4185 		if (t < 0 || t >= NBGE_RX_THRESH)
4186 			return (EINVAL);
4187 		bge_update_all_threshes(t);
4188 	} else
4189 		return (EINVAL);
4190 
4191 	*(int*)rnode->sysctl_data = t;
4192 
4193 	return (0);
4194 }
4195 
4196 /*
4197  * Set up sysctl(3) MIB, hw.bge.*.
4198  *
4199  * TBD condition SYSCTL_PERMANENT on being an LKM or not
4200  */
4201 SYSCTL_SETUP(sysctl_bge, "sysctl bge subtree setup")
4202 {
4203 	int rc, bge_root_num;
4204 	const struct sysctlnode *node;
4205 
4206 	if ((rc = sysctl_createv(clog, 0, NULL, NULL,
4207 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "hw", NULL,
4208 	    NULL, 0, NULL, 0, CTL_HW, CTL_EOL)) != 0) {
4209 		goto err;
4210 	}
4211 
4212 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
4213 	    CTLFLAG_PERMANENT, CTLTYPE_NODE, "bge",
4214 	    SYSCTL_DESCR("BGE interface controls"),
4215 	    NULL, 0, NULL, 0, CTL_HW, CTL_CREATE, CTL_EOL)) != 0) {
4216 		goto err;
4217 	}
4218 
4219 	bge_root_num = node->sysctl_num;
4220 
4221 	/* BGE Rx interrupt mitigation level */
4222 	if ((rc = sysctl_createv(clog, 0, NULL, &node,
4223 	    CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
4224 	    CTLTYPE_INT, "rx_lvl",
4225 	    SYSCTL_DESCR("BGE receive interrupt mitigation level"),
4226 	    sysctl_bge_verify, 0,
4227 	    &bge_rx_thresh_lvl,
4228 	    0, CTL_HW, bge_root_num, CTL_CREATE,
4229 	    CTL_EOL)) != 0) {
4230 		goto err;
4231 	}
4232 
4233 	bge_rxthresh_nodenum = node->sysctl_num;
4234 
4235 	return;
4236 
4237 err:
4238 	printf("%s: sysctl_createv failed (rc = %d)\n", __func__, rc);
4239 }
4240 
4241 static void
4242 bge_powerhook(int why, void *hdl)
4243 {
4244 	struct bge_softc *sc = (struct bge_softc *)hdl;
4245 	struct ifnet *ifp = &sc->ethercom.ec_if;
4246 	struct pci_attach_args *pa = &(sc->bge_pa);
4247 	pci_chipset_tag_t pc = pa->pa_pc;
4248 	pcitag_t tag = pa->pa_tag;
4249 
4250 	switch (why) {
4251 	case PWR_SOFTSUSPEND:
4252 	case PWR_SOFTSTANDBY:
4253 		bge_shutdown(sc);
4254 		break;
4255 	case PWR_SOFTRESUME:
4256 		if (ifp->if_flags & IFF_UP) {
4257 			ifp->if_flags &= ~IFF_RUNNING;
4258 			bge_init(ifp);
4259 		}
4260 		break;
4261 	case PWR_SUSPEND:
4262 	case PWR_STANDBY:
4263 		pci_conf_capture(pc, tag, &sc->bge_pciconf);
4264 		break;
4265 	case PWR_RESUME:
4266 		pci_conf_restore(pc, tag, &sc->bge_pciconf);
4267 		break;
4268 	}
4269 
4270 	return;
4271 }
4272