xref: /netbsd-src/sys/dev/pci/if_vge.c (revision 404fbe5fb94ca1e054339640cabb2801ce52dd30)
1 /* $NetBSD: if_vge.c,v 1.43 2008/11/26 21:34:07 joerg Exp $ */
2 
3 /*-
4  * Copyright (c) 2004
5  *	Bill Paul <wpaul@windriver.com>.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. All advertising materials mentioning features or use of this software
16  *    must display the following acknowledgement:
17  *	This product includes software developed by Bill Paul.
18  * 4. Neither the name of the author nor the names of any co-contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY Bill Paul AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL Bill Paul OR THE VOICES IN HIS HEAD
26  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
27  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
28  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
29  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
30  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
31  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
32  * THE POSSIBILITY OF SUCH DAMAGE.
33  *
34  * FreeBSD: src/sys/dev/vge/if_vge.c,v 1.5 2005/02/07 19:39:29 glebius Exp
35  */
36 
37 #include <sys/cdefs.h>
38 __KERNEL_RCSID(0, "$NetBSD: if_vge.c,v 1.43 2008/11/26 21:34:07 joerg Exp $");
39 
40 /*
41  * VIA Networking Technologies VT612x PCI gigabit ethernet NIC driver.
42  *
43  * Written by Bill Paul <wpaul@windriver.com>
44  * Senior Networking Software Engineer
45  * Wind River Systems
46  */
47 
48 /*
49  * The VIA Networking VT6122 is a 32bit, 33/66 MHz PCI device that
50  * combines a tri-speed ethernet MAC and PHY, with the following
51  * features:
52  *
53  *	o Jumbo frame support up to 16K
54  *	o Transmit and receive flow control
55  *	o IPv4 checksum offload
56  *	o VLAN tag insertion and stripping
57  *	o TCP large send
58  *	o 64-bit multicast hash table filter
59  *	o 64 entry CAM filter
60  *	o 16K RX FIFO and 48K TX FIFO memory
61  *	o Interrupt moderation
62  *
63  * The VT6122 supports up to four transmit DMA queues. The descriptors
64  * in the transmit ring can address up to 7 data fragments; frames which
65  * span more than 7 data buffers must be coalesced, but in general the
66  * BSD TCP/IP stack rarely generates frames more than 2 or 3 fragments
67  * long. The receive descriptors address only a single buffer.
68  *
69  * There are two peculiar design issues with the VT6122. One is that
70  * receive data buffers must be aligned on a 32-bit boundary. This is
71  * not a problem where the VT6122 is used as a LOM device in x86-based
72  * systems, but on architectures that generate unaligned access traps, we
73  * have to do some copying.
74  *
75  * The other issue has to do with the way 64-bit addresses are handled.
76  * The DMA descriptors only allow you to specify 48 bits of addressing
77  * information. The remaining 16 bits are specified using one of the
78  * I/O registers. If you only have a 32-bit system, then this isn't
79  * an issue, but if you have a 64-bit system and more than 4GB of
80  * memory, you must have to make sure your network data buffers reside
81  * in the same 48-bit 'segment.'
82  *
83  * Special thanks to Ryan Fu at VIA Networking for providing documentation
84  * and sample NICs for testing.
85  */
86 
87 #include "bpfilter.h"
88 
89 #include <sys/param.h>
90 #include <sys/endian.h>
91 #include <sys/systm.h>
92 #include <sys/device.h>
93 #include <sys/sockio.h>
94 #include <sys/mbuf.h>
95 #include <sys/malloc.h>
96 #include <sys/kernel.h>
97 #include <sys/socket.h>
98 
99 #include <net/if.h>
100 #include <net/if_arp.h>
101 #include <net/if_ether.h>
102 #include <net/if_dl.h>
103 #include <net/if_media.h>
104 
105 #include <net/bpf.h>
106 
107 #include <sys/bus.h>
108 
109 #include <dev/mii/mii.h>
110 #include <dev/mii/miivar.h>
111 
112 #include <dev/pci/pcireg.h>
113 #include <dev/pci/pcivar.h>
114 #include <dev/pci/pcidevs.h>
115 
116 #include <dev/pci/if_vgereg.h>
117 
118 #define VGE_IFQ_MAXLEN		64
119 
120 #define VGE_RING_ALIGN		256
121 
122 #define VGE_NTXDESC		256
123 #define VGE_NTXDESC_MASK	(VGE_NTXDESC - 1)
124 #define VGE_NEXT_TXDESC(x)	((x + 1) & VGE_NTXDESC_MASK)
125 #define VGE_PREV_TXDESC(x)	((x - 1) & VGE_NTXDESC_MASK)
126 
127 #define VGE_NRXDESC		256	/* Must be a multiple of 4!! */
128 #define VGE_NRXDESC_MASK	(VGE_NRXDESC - 1)
129 #define VGE_NEXT_RXDESC(x)	((x + 1) & VGE_NRXDESC_MASK)
130 #define VGE_PREV_RXDESC(x)	((x - 1) & VGE_NRXDESC_MASK)
131 
132 #define VGE_ADDR_LO(y)		((uint64_t)(y) & 0xFFFFFFFF)
133 #define VGE_ADDR_HI(y)		((uint64_t)(y) >> 32)
134 #define VGE_BUFLEN(y)		((y) & 0x7FFF)
135 #define ETHER_PAD_LEN		(ETHER_MIN_LEN - ETHER_CRC_LEN)
136 
137 #define VGE_POWER_MANAGEMENT	0	/* disabled for now */
138 
139 /*
140  * Mbuf adjust factor to force 32-bit alignment of IP header.
141  * Drivers should pad ETHER_ALIGN bytes when setting up a
142  * RX mbuf so the upper layers get the IP header properly aligned
143  * past the 14-byte Ethernet header.
144  *
145  * See also comment in vge_encap().
146  */
147 #define ETHER_ALIGN		2
148 
149 #ifdef __NO_STRICT_ALIGNMENT
150 #define VGE_RX_BUFSIZE		MCLBYTES
151 #else
152 #define VGE_RX_PAD		sizeof(uint32_t)
153 #define VGE_RX_BUFSIZE		(MCLBYTES - VGE_RX_PAD)
154 #endif
155 
156 /*
157  * Control structures are DMA'd to the vge chip. We allocate them in
158  * a single clump that maps to a single DMA segment to make several things
159  * easier.
160  */
161 struct vge_control_data {
162 	/* TX descriptors */
163 	struct vge_txdesc	vcd_txdescs[VGE_NTXDESC];
164 	/* RX descriptors */
165 	struct vge_rxdesc	vcd_rxdescs[VGE_NRXDESC];
166 	/* dummy data for TX padding */
167 	uint8_t			vcd_pad[ETHER_PAD_LEN];
168 };
169 
170 #define VGE_CDOFF(x)	offsetof(struct vge_control_data, x)
171 #define VGE_CDTXOFF(x)	VGE_CDOFF(vcd_txdescs[(x)])
172 #define VGE_CDRXOFF(x)	VGE_CDOFF(vcd_rxdescs[(x)])
173 #define VGE_CDPADOFF()	VGE_CDOFF(vcd_pad[0])
174 
175 /*
176  * Software state for TX jobs.
177  */
178 struct vge_txsoft {
179 	struct mbuf	*txs_mbuf;		/* head of our mbuf chain */
180 	bus_dmamap_t	txs_dmamap;		/* our DMA map */
181 };
182 
183 /*
184  * Software state for RX jobs.
185  */
186 struct vge_rxsoft {
187 	struct mbuf	*rxs_mbuf;		/* head of our mbuf chain */
188 	bus_dmamap_t	rxs_dmamap;		/* our DMA map */
189 };
190 
191 
192 struct vge_softc {
193 	struct device		sc_dev;
194 
195 	bus_space_tag_t		sc_bst;		/* bus space tag */
196 	bus_space_handle_t	sc_bsh;		/* bus space handle */
197 	bus_dma_tag_t		sc_dmat;
198 
199 	struct ethercom		sc_ethercom;	/* interface info */
200 	uint8_t			sc_eaddr[ETHER_ADDR_LEN];
201 
202 	void			*sc_intrhand;
203 	struct mii_data		sc_mii;
204 	uint8_t			sc_type;
205 	int			sc_if_flags;
206 	int			sc_link;
207 	int			sc_camidx;
208 	callout_t		sc_timeout;
209 
210 	bus_dmamap_t		sc_cddmamap;
211 #define sc_cddma		sc_cddmamap->dm_segs[0].ds_addr
212 
213 	struct vge_txsoft	sc_txsoft[VGE_NTXDESC];
214 	struct vge_rxsoft	sc_rxsoft[VGE_NRXDESC];
215 	struct vge_control_data	*sc_control_data;
216 #define sc_txdescs		sc_control_data->vcd_txdescs
217 #define sc_rxdescs		sc_control_data->vcd_rxdescs
218 
219 	int			sc_tx_prodidx;
220 	int			sc_tx_considx;
221 	int			sc_tx_free;
222 
223 	struct mbuf		*sc_rx_mhead;
224 	struct mbuf		*sc_rx_mtail;
225 	int			sc_rx_prodidx;
226 	int			sc_rx_consumed;
227 
228 	int			sc_suspended;	/* 0 = normal  1 = suspended */
229 	uint32_t		sc_saved_maps[5];	/* pci data */
230 	uint32_t		sc_saved_biosaddr;
231 	uint8_t			sc_saved_intline;
232 	uint8_t			sc_saved_cachelnsz;
233 	uint8_t			sc_saved_lattimer;
234 };
235 
236 #define VGE_CDTXADDR(sc, x)	((sc)->sc_cddma + VGE_CDTXOFF(x))
237 #define VGE_CDRXADDR(sc, x)	((sc)->sc_cddma + VGE_CDRXOFF(x))
238 #define VGE_CDPADADDR(sc)	((sc)->sc_cddma + VGE_CDPADOFF())
239 
240 #define VGE_TXDESCSYNC(sc, idx, ops)					\
241 	bus_dmamap_sync((sc)->sc_dmat,(sc)->sc_cddmamap,		\
242 	    VGE_CDTXOFF(idx),						\
243 	    offsetof(struct vge_txdesc, td_frag[0]),			\
244 	    (ops))
245 #define VGE_TXFRAGSYNC(sc, idx, nsegs, ops)				\
246 	bus_dmamap_sync((sc)->sc_dmat, (sc)->sc_cddmamap,		\
247 	    VGE_CDTXOFF(idx) +						\
248 	    offsetof(struct vge_txdesc, td_frag[0]),			\
249 	    sizeof(struct vge_txfrag) * (nsegs),			\
250 	    (ops))
251 #define VGE_RXDESCSYNC(sc, idx, ops)					\
252 	bus_dmamap_sync((sc)->sc_dmat, (sc)->sc_cddmamap,		\
253 	    VGE_CDRXOFF(idx),						\
254 	    sizeof(struct vge_rxdesc),					\
255 	    (ops))
256 
257 /*
258  * register space access macros
259  */
260 #define CSR_WRITE_4(sc, reg, val)	\
261 	bus_space_write_4((sc)->sc_bst, (sc)->sc_bsh, (reg), (val))
262 #define CSR_WRITE_2(sc, reg, val)	\
263 	bus_space_write_2((sc)->sc_bst, (sc)->sc_bsh, (reg), (val))
264 #define CSR_WRITE_1(sc, reg, val)	\
265 	bus_space_write_1((sc)->sc_bst, (sc)->sc_bsh, (reg), (val))
266 
267 #define CSR_READ_4(sc, reg)		\
268 	bus_space_read_4((sc)->sc_bst, (sc)->sc_bsh, (reg))
269 #define CSR_READ_2(sc, reg)		\
270 	bus_space_read_2((sc)->sc_bst, (sc)->sc_bsh, (reg))
271 #define CSR_READ_1(sc, reg)		\
272 	bus_space_read_1((sc)->sc_bst, (sc)->sc_bsh, (reg))
273 
274 #define CSR_SETBIT_1(sc, reg, x)	\
275 	CSR_WRITE_1((sc), (reg), CSR_READ_1((sc), (reg)) | (x))
276 #define CSR_SETBIT_2(sc, reg, x)	\
277 	CSR_WRITE_2((sc), (reg), CSR_READ_2((sc), (reg)) | (x))
278 #define CSR_SETBIT_4(sc, reg, x)	\
279 	CSR_WRITE_4((sc), (reg), CSR_READ_4((sc), (reg)) | (x))
280 
281 #define CSR_CLRBIT_1(sc, reg, x)	\
282 	CSR_WRITE_1((sc), (reg), CSR_READ_1((sc), (reg)) & ~(x))
283 #define CSR_CLRBIT_2(sc, reg, x)	\
284 	CSR_WRITE_2((sc), (reg), CSR_READ_2((sc), (reg)) & ~(x))
285 #define CSR_CLRBIT_4(sc, reg, x)	\
286 	CSR_WRITE_4((sc), (reg), CSR_READ_4((sc), (reg)) & ~(x))
287 
288 #define VGE_TIMEOUT		10000
289 
290 #define VGE_PCI_LOIO             0x10
291 #define VGE_PCI_LOMEM            0x14
292 
293 static inline void vge_set_txaddr(struct vge_txfrag *, bus_addr_t);
294 static inline void vge_set_rxaddr(struct vge_rxdesc *, bus_addr_t);
295 
296 static int vge_ifflags_cb(struct ethercom *);
297 
298 static int vge_match(struct device *, struct cfdata *, void *);
299 static void vge_attach(struct device *, struct device *, void *);
300 
301 static int vge_encap(struct vge_softc *, struct mbuf *, int);
302 
303 static int vge_allocmem(struct vge_softc *);
304 static int vge_newbuf(struct vge_softc *, int, struct mbuf *);
305 #ifndef __NO_STRICT_ALIGNMENT
306 static inline void vge_fixup_rx(struct mbuf *);
307 #endif
308 static void vge_rxeof(struct vge_softc *);
309 static void vge_txeof(struct vge_softc *);
310 static int vge_intr(void *);
311 static void vge_tick(void *);
312 static void vge_start(struct ifnet *);
313 static int vge_ioctl(struct ifnet *, u_long, void *);
314 static int vge_init(struct ifnet *);
315 static void vge_stop(struct ifnet *, int);
316 static void vge_watchdog(struct ifnet *);
317 #if VGE_POWER_MANAGEMENT
318 static int vge_suspend(struct device *);
319 static int vge_resume(struct device *);
320 #endif
321 static void vge_shutdown(void *);
322 
323 static uint16_t vge_read_eeprom(struct vge_softc *, int);
324 
325 static void vge_miipoll_start(struct vge_softc *);
326 static void vge_miipoll_stop(struct vge_softc *);
327 static int vge_miibus_readreg(struct device *, int, int);
328 static void vge_miibus_writereg(struct device *, int, int, int);
329 static void vge_miibus_statchg(struct device *);
330 
331 static void vge_cam_clear(struct vge_softc *);
332 static int vge_cam_set(struct vge_softc *, uint8_t *);
333 static void vge_setmulti(struct vge_softc *);
334 static void vge_reset(struct vge_softc *);
335 
336 CFATTACH_DECL(vge, sizeof(struct vge_softc),
337     vge_match, vge_attach, NULL, NULL);
338 
339 static inline void
340 vge_set_txaddr(struct vge_txfrag *f, bus_addr_t daddr)
341 {
342 
343 	f->tf_addrlo = htole32((uint32_t)daddr);
344 	if (sizeof(bus_addr_t) == sizeof(uint64_t))
345 		f->tf_addrhi = htole16(((uint64_t)daddr >> 32) & 0xFFFF);
346 	else
347 		f->tf_addrhi = 0;
348 }
349 
350 static inline void
351 vge_set_rxaddr(struct vge_rxdesc *rxd, bus_addr_t daddr)
352 {
353 
354 	rxd->rd_addrlo = htole32((uint32_t)daddr);
355 	if (sizeof(bus_addr_t) == sizeof(uint64_t))
356 		rxd->rd_addrhi = htole16(((uint64_t)daddr >> 32) & 0xFFFF);
357 	else
358 		rxd->rd_addrhi = 0;
359 }
360 
361 /*
362  * Defragment mbuf chain contents to be as linear as possible.
363  * Returns new mbuf chain on success, NULL on failure. Old mbuf
364  * chain is always freed.
365  * XXX temporary until there would be generic function doing this.
366  */
367 #define m_defrag	vge_m_defrag
368 struct mbuf * vge_m_defrag(struct mbuf *, int);
369 
370 struct mbuf *
371 vge_m_defrag(struct mbuf *mold, int flags)
372 {
373 	struct mbuf *m0, *mn, *n;
374 	size_t sz = mold->m_pkthdr.len;
375 
376 #ifdef DIAGNOSTIC
377 	if ((mold->m_flags & M_PKTHDR) == 0)
378 		panic("m_defrag: not a mbuf chain header");
379 #endif
380 
381 	MGETHDR(m0, flags, MT_DATA);
382 	if (m0 == NULL)
383 		return NULL;
384 	m0->m_pkthdr.len = mold->m_pkthdr.len;
385 	mn = m0;
386 
387 	do {
388 		if (sz > MHLEN) {
389 			MCLGET(mn, M_DONTWAIT);
390 			if ((mn->m_flags & M_EXT) == 0) {
391 				m_freem(m0);
392 				return NULL;
393 			}
394 		}
395 
396 		mn->m_len = MIN(sz, MCLBYTES);
397 
398 		m_copydata(mold, mold->m_pkthdr.len - sz, mn->m_len,
399 		     mtod(mn, void *));
400 
401 		sz -= mn->m_len;
402 
403 		if (sz > 0) {
404 			/* need more mbufs */
405 			MGET(n, M_NOWAIT, MT_DATA);
406 			if (n == NULL) {
407 				m_freem(m0);
408 				return NULL;
409 			}
410 
411 			mn->m_next = n;
412 			mn = n;
413 		}
414 	} while (sz > 0);
415 
416 	return m0;
417 }
418 
419 /*
420  * Read a word of data stored in the EEPROM at address 'addr.'
421  */
422 static uint16_t
423 vge_read_eeprom(struct vge_softc *sc, int addr)
424 {
425 	int i;
426 	uint16_t word = 0;
427 
428 	/*
429 	 * Enter EEPROM embedded programming mode. In order to
430 	 * access the EEPROM at all, we first have to set the
431 	 * EELOAD bit in the CHIPCFG2 register.
432 	 */
433 	CSR_SETBIT_1(sc, VGE_CHIPCFG2, VGE_CHIPCFG2_EELOAD);
434 	CSR_SETBIT_1(sc, VGE_EECSR, VGE_EECSR_EMBP/*|VGE_EECSR_ECS*/);
435 
436 	/* Select the address of the word we want to read */
437 	CSR_WRITE_1(sc, VGE_EEADDR, addr);
438 
439 	/* Issue read command */
440 	CSR_SETBIT_1(sc, VGE_EECMD, VGE_EECMD_ERD);
441 
442 	/* Wait for the done bit to be set. */
443 	for (i = 0; i < VGE_TIMEOUT; i++) {
444 		if (CSR_READ_1(sc, VGE_EECMD) & VGE_EECMD_EDONE)
445 			break;
446 	}
447 
448 	if (i == VGE_TIMEOUT) {
449 		aprint_error_dev(&sc->sc_dev, "EEPROM read timed out\n");
450 		return 0;
451 	}
452 
453 	/* Read the result */
454 	word = CSR_READ_2(sc, VGE_EERDDAT);
455 
456 	/* Turn off EEPROM access mode. */
457 	CSR_CLRBIT_1(sc, VGE_EECSR, VGE_EECSR_EMBP/*|VGE_EECSR_ECS*/);
458 	CSR_CLRBIT_1(sc, VGE_CHIPCFG2, VGE_CHIPCFG2_EELOAD);
459 
460 	return word;
461 }
462 
463 static void
464 vge_miipoll_stop(struct vge_softc *sc)
465 {
466 	int i;
467 
468 	CSR_WRITE_1(sc, VGE_MIICMD, 0);
469 
470 	for (i = 0; i < VGE_TIMEOUT; i++) {
471 		DELAY(1);
472 		if (CSR_READ_1(sc, VGE_MIISTS) & VGE_MIISTS_IIDL)
473 			break;
474 	}
475 
476 	if (i == VGE_TIMEOUT) {
477 		aprint_error_dev(&sc->sc_dev, "failed to idle MII autopoll\n");
478 	}
479 }
480 
481 static void
482 vge_miipoll_start(struct vge_softc *sc)
483 {
484 	int i;
485 
486 	/* First, make sure we're idle. */
487 
488 	CSR_WRITE_1(sc, VGE_MIICMD, 0);
489 	CSR_WRITE_1(sc, VGE_MIIADDR, VGE_MIIADDR_SWMPL);
490 
491 	for (i = 0; i < VGE_TIMEOUT; i++) {
492 		DELAY(1);
493 		if (CSR_READ_1(sc, VGE_MIISTS) & VGE_MIISTS_IIDL)
494 			break;
495 	}
496 
497 	if (i == VGE_TIMEOUT) {
498 		aprint_error_dev(&sc->sc_dev, "failed to idle MII autopoll\n");
499 		return;
500 	}
501 
502 	/* Now enable auto poll mode. */
503 
504 	CSR_WRITE_1(sc, VGE_MIICMD, VGE_MIICMD_MAUTO);
505 
506 	/* And make sure it started. */
507 
508 	for (i = 0; i < VGE_TIMEOUT; i++) {
509 		DELAY(1);
510 		if ((CSR_READ_1(sc, VGE_MIISTS) & VGE_MIISTS_IIDL) == 0)
511 			break;
512 	}
513 
514 	if (i == VGE_TIMEOUT) {
515 		aprint_error_dev(&sc->sc_dev, "failed to start MII autopoll\n");
516 	}
517 }
518 
519 static int
520 vge_miibus_readreg(struct device *dev, int phy, int reg)
521 {
522 	struct vge_softc *sc;
523 	int i, s;
524 	uint16_t rval;
525 
526 	sc = (void *)dev;
527 	rval = 0;
528 	if (phy != (CSR_READ_1(sc, VGE_MIICFG) & 0x1F))
529 		return 0;
530 
531 	s = splnet();
532 	vge_miipoll_stop(sc);
533 
534 	/* Specify the register we want to read. */
535 	CSR_WRITE_1(sc, VGE_MIIADDR, reg);
536 
537 	/* Issue read command. */
538 	CSR_SETBIT_1(sc, VGE_MIICMD, VGE_MIICMD_RCMD);
539 
540 	/* Wait for the read command bit to self-clear. */
541 	for (i = 0; i < VGE_TIMEOUT; i++) {
542 		DELAY(1);
543 		if ((CSR_READ_1(sc, VGE_MIICMD) & VGE_MIICMD_RCMD) == 0)
544 			break;
545 	}
546 
547 	if (i == VGE_TIMEOUT)
548 		aprint_error_dev(&sc->sc_dev, "MII read timed out\n");
549 	else
550 		rval = CSR_READ_2(sc, VGE_MIIDATA);
551 
552 	vge_miipoll_start(sc);
553 	splx(s);
554 
555 	return rval;
556 }
557 
558 static void
559 vge_miibus_writereg(struct device *dev, int phy, int reg, int data)
560 {
561 	struct vge_softc *sc;
562 	int i, s;
563 
564 	sc = (void *)dev;
565 	if (phy != (CSR_READ_1(sc, VGE_MIICFG) & 0x1F))
566 		return;
567 
568 	s = splnet();
569 	vge_miipoll_stop(sc);
570 
571 	/* Specify the register we want to write. */
572 	CSR_WRITE_1(sc, VGE_MIIADDR, reg);
573 
574 	/* Specify the data we want to write. */
575 	CSR_WRITE_2(sc, VGE_MIIDATA, data);
576 
577 	/* Issue write command. */
578 	CSR_SETBIT_1(sc, VGE_MIICMD, VGE_MIICMD_WCMD);
579 
580 	/* Wait for the write command bit to self-clear. */
581 	for (i = 0; i < VGE_TIMEOUT; i++) {
582 		DELAY(1);
583 		if ((CSR_READ_1(sc, VGE_MIICMD) & VGE_MIICMD_WCMD) == 0)
584 			break;
585 	}
586 
587 	if (i == VGE_TIMEOUT) {
588 		aprint_error_dev(&sc->sc_dev, "MII write timed out\n");
589 	}
590 
591 	vge_miipoll_start(sc);
592 	splx(s);
593 }
594 
595 static void
596 vge_cam_clear(struct vge_softc *sc)
597 {
598 	int i;
599 
600 	/*
601 	 * Turn off all the mask bits. This tells the chip
602 	 * that none of the entries in the CAM filter are valid.
603 	 * desired entries will be enabled as we fill the filter in.
604 	 */
605 
606 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
607 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_CAMMASK);
608 	CSR_WRITE_1(sc, VGE_CAMADDR, VGE_CAMADDR_ENABLE);
609 	for (i = 0; i < 8; i++)
610 		CSR_WRITE_1(sc, VGE_CAM0 + i, 0);
611 
612 	/* Clear the VLAN filter too. */
613 
614 	CSR_WRITE_1(sc, VGE_CAMADDR, VGE_CAMADDR_ENABLE|VGE_CAMADDR_AVSEL|0);
615 	for (i = 0; i < 8; i++)
616 		CSR_WRITE_1(sc, VGE_CAM0 + i, 0);
617 
618 	CSR_WRITE_1(sc, VGE_CAMADDR, 0);
619 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
620 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_MAR);
621 
622 	sc->sc_camidx = 0;
623 }
624 
625 static int
626 vge_cam_set(struct vge_softc *sc, uint8_t *addr)
627 {
628 	int i, error;
629 
630 	error = 0;
631 
632 	if (sc->sc_camidx == VGE_CAM_MAXADDRS)
633 		return ENOSPC;
634 
635 	/* Select the CAM data page. */
636 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
637 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_CAMDATA);
638 
639 	/* Set the filter entry we want to update and enable writing. */
640 	CSR_WRITE_1(sc, VGE_CAMADDR, VGE_CAMADDR_ENABLE | sc->sc_camidx);
641 
642 	/* Write the address to the CAM registers */
643 	for (i = 0; i < ETHER_ADDR_LEN; i++)
644 		CSR_WRITE_1(sc, VGE_CAM0 + i, addr[i]);
645 
646 	/* Issue a write command. */
647 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_WRITE);
648 
649 	/* Wake for it to clear. */
650 	for (i = 0; i < VGE_TIMEOUT; i++) {
651 		DELAY(1);
652 		if ((CSR_READ_1(sc, VGE_CAMCTL) & VGE_CAMCTL_WRITE) == 0)
653 			break;
654 	}
655 
656 	if (i == VGE_TIMEOUT) {
657 		aprint_error_dev(&sc->sc_dev, "setting CAM filter failed\n");
658 		error = EIO;
659 		goto fail;
660 	}
661 
662 	/* Select the CAM mask page. */
663 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
664 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_CAMMASK);
665 
666 	/* Set the mask bit that enables this filter. */
667 	CSR_SETBIT_1(sc, VGE_CAM0 + (sc->sc_camidx / 8),
668 	    1 << (sc->sc_camidx & 7));
669 
670 	sc->sc_camidx++;
671 
672  fail:
673 	/* Turn off access to CAM. */
674 	CSR_WRITE_1(sc, VGE_CAMADDR, 0);
675 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
676 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_MAR);
677 
678 	return error;
679 }
680 
681 /*
682  * Program the multicast filter. We use the 64-entry CAM filter
683  * for perfect filtering. If there's more than 64 multicast addresses,
684  * we use the hash filter instead.
685  */
686 static void
687 vge_setmulti(struct vge_softc *sc)
688 {
689 	struct ifnet *ifp;
690 	int error;
691 	uint32_t h, hashes[2] = { 0, 0 };
692 	struct ether_multi *enm;
693 	struct ether_multistep step;
694 
695 	error = 0;
696 	ifp = &sc->sc_ethercom.ec_if;
697 
698 	/* First, zot all the multicast entries. */
699 	vge_cam_clear(sc);
700 	CSR_WRITE_4(sc, VGE_MAR0, 0);
701 	CSR_WRITE_4(sc, VGE_MAR1, 0);
702 	ifp->if_flags &= ~IFF_ALLMULTI;
703 
704 	/*
705 	 * If the user wants allmulti or promisc mode, enable reception
706 	 * of all multicast frames.
707 	 */
708 	if (ifp->if_flags & IFF_PROMISC) {
709  allmulti:
710 		CSR_WRITE_4(sc, VGE_MAR0, 0xFFFFFFFF);
711 		CSR_WRITE_4(sc, VGE_MAR1, 0xFFFFFFFF);
712 		ifp->if_flags |= IFF_ALLMULTI;
713 		return;
714 	}
715 
716 	/* Now program new ones */
717 	ETHER_FIRST_MULTI(step, &sc->sc_ethercom, enm);
718 	while (enm != NULL) {
719 		/*
720 		 * If multicast range, fall back to ALLMULTI.
721 		 */
722 		if (memcmp(enm->enm_addrlo, enm->enm_addrhi,
723 				ETHER_ADDR_LEN) != 0)
724 			goto allmulti;
725 
726 		error = vge_cam_set(sc, enm->enm_addrlo);
727 		if (error)
728 			break;
729 
730 		ETHER_NEXT_MULTI(step, enm);
731 	}
732 
733 	/* If there were too many addresses, use the hash filter. */
734 	if (error) {
735 		vge_cam_clear(sc);
736 
737 		ETHER_FIRST_MULTI(step, &sc->sc_ethercom, enm);
738 		while (enm != NULL) {
739 			/*
740 			 * If multicast range, fall back to ALLMULTI.
741 			 */
742 			if (memcmp(enm->enm_addrlo, enm->enm_addrhi,
743 					ETHER_ADDR_LEN) != 0)
744 				goto allmulti;
745 
746 			h = ether_crc32_be(enm->enm_addrlo,
747 			    ETHER_ADDR_LEN) >> 26;
748 			hashes[h >> 5] |= 1 << (h & 0x1f);
749 
750 			ETHER_NEXT_MULTI(step, enm);
751 		}
752 
753 		CSR_WRITE_4(sc, VGE_MAR0, hashes[0]);
754 		CSR_WRITE_4(sc, VGE_MAR1, hashes[1]);
755 	}
756 }
757 
758 static void
759 vge_reset(struct vge_softc *sc)
760 {
761 	int i;
762 
763 	CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_SOFTRESET);
764 
765 	for (i = 0; i < VGE_TIMEOUT; i++) {
766 		DELAY(5);
767 		if ((CSR_READ_1(sc, VGE_CRS1) & VGE_CR1_SOFTRESET) == 0)
768 			break;
769 	}
770 
771 	if (i == VGE_TIMEOUT) {
772 		aprint_error_dev(&sc->sc_dev, "soft reset timed out");
773 		CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_STOP_FORCE);
774 		DELAY(2000);
775 	}
776 
777 	DELAY(5000);
778 
779 	CSR_SETBIT_1(sc, VGE_EECSR, VGE_EECSR_RELOAD);
780 
781 	for (i = 0; i < VGE_TIMEOUT; i++) {
782 		DELAY(5);
783 		if ((CSR_READ_1(sc, VGE_EECSR) & VGE_EECSR_RELOAD) == 0)
784 			break;
785 	}
786 
787 	if (i == VGE_TIMEOUT) {
788 		aprint_error_dev(&sc->sc_dev, "EEPROM reload timed out\n");
789 		return;
790 	}
791 
792 	/*
793 	 * On some machine, the first read data from EEPROM could be
794 	 * messed up, so read one dummy data here to avoid the mess.
795 	 */
796 	(void)vge_read_eeprom(sc, 0);
797 
798 	CSR_CLRBIT_1(sc, VGE_CHIPCFG0, VGE_CHIPCFG0_PACPI);
799 }
800 
801 /*
802  * Probe for a VIA gigabit chip. Check the PCI vendor and device
803  * IDs against our list and return a device name if we find a match.
804  */
805 static int
806 vge_match(struct device *parent, struct cfdata *match, void *aux)
807 {
808 	struct pci_attach_args *pa = aux;
809 
810 	if (PCI_VENDOR(pa->pa_id) == PCI_VENDOR_VIATECH
811 	    && PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_VIATECH_VT612X)
812 		return 1;
813 
814 	return 0;
815 }
816 
817 static int
818 vge_allocmem(struct vge_softc *sc)
819 {
820 	int error;
821 	int nseg;
822 	int i;
823 	bus_dma_segment_t seg;
824 
825 	/*
826 	 * Allocate memory for control data.
827 	 */
828 
829 	error = bus_dmamem_alloc(sc->sc_dmat, sizeof(struct vge_control_data),
830 	     VGE_RING_ALIGN, 0, &seg, 1, &nseg, BUS_DMA_NOWAIT);
831 	if (error) {
832 		aprint_error_dev(&sc->sc_dev, "could not allocate control data dma memory\n");
833 		goto fail_1;
834 	}
835 
836 	/* Map the memory to kernel VA space */
837 
838 	error = bus_dmamem_map(sc->sc_dmat, &seg, nseg,
839 	    sizeof(struct vge_control_data), (void **)&sc->sc_control_data,
840 	    BUS_DMA_NOWAIT);
841 	if (error) {
842 		aprint_error_dev(&sc->sc_dev, "could not map control data dma memory\n");
843 		goto fail_2;
844 	}
845 	memset(sc->sc_control_data, 0, sizeof(struct vge_control_data));
846 
847 	/*
848 	 * Create map for control data.
849 	 */
850 	error = bus_dmamap_create(sc->sc_dmat,
851 	    sizeof(struct vge_control_data), 1,
852 	    sizeof(struct vge_control_data), 0, BUS_DMA_NOWAIT,
853 	    &sc->sc_cddmamap);
854 	if (error) {
855 		aprint_error_dev(&sc->sc_dev, "could not create control data dmamap\n");
856 		goto fail_3;
857 	}
858 
859 	/* Load the map for the control data. */
860 	error = bus_dmamap_load(sc->sc_dmat, sc->sc_cddmamap,
861 	    sc->sc_control_data, sizeof(struct vge_control_data), NULL,
862 	    BUS_DMA_NOWAIT);
863 	if (error) {
864 		aprint_error_dev(&sc->sc_dev, "could not load control data dma memory\n");
865 		goto fail_4;
866 	}
867 
868 	/* Create DMA maps for TX buffers */
869 
870 	for (i = 0; i < VGE_NTXDESC; i++) {
871 		error = bus_dmamap_create(sc->sc_dmat, VGE_TX_MAXLEN,
872 		    VGE_TX_FRAGS, VGE_TX_MAXLEN, 0, BUS_DMA_NOWAIT,
873 		    &sc->sc_txsoft[i].txs_dmamap);
874 		if (error) {
875 			aprint_error_dev(&sc->sc_dev, "can't create DMA map for TX descs\n");
876 			goto fail_5;
877 		}
878 	}
879 
880 	/* Create DMA maps for RX buffers */
881 
882 	for (i = 0; i < VGE_NRXDESC; i++) {
883 		error = bus_dmamap_create(sc->sc_dmat, MCLBYTES,
884 		    1, MCLBYTES, 0, BUS_DMA_NOWAIT,
885 		    &sc->sc_rxsoft[i].rxs_dmamap);
886 		if (error) {
887 			aprint_error_dev(&sc->sc_dev, "can't create DMA map for RX descs\n");
888 			goto fail_6;
889 		}
890 		sc->sc_rxsoft[i].rxs_mbuf = NULL;
891 	}
892 
893 	return 0;
894 
895  fail_6:
896 	for (i = 0; i < VGE_NRXDESC; i++) {
897 		if (sc->sc_rxsoft[i].rxs_dmamap != NULL)
898 			bus_dmamap_destroy(sc->sc_dmat,
899 			    sc->sc_rxsoft[i].rxs_dmamap);
900 	}
901  fail_5:
902 	for (i = 0; i < VGE_NTXDESC; i++) {
903 		if (sc->sc_txsoft[i].txs_dmamap != NULL)
904 			bus_dmamap_destroy(sc->sc_dmat,
905 			    sc->sc_txsoft[i].txs_dmamap);
906 	}
907 	bus_dmamap_unload(sc->sc_dmat, sc->sc_cddmamap);
908  fail_4:
909 	bus_dmamap_destroy(sc->sc_dmat, sc->sc_cddmamap);
910  fail_3:
911 	bus_dmamem_unmap(sc->sc_dmat, (void *)sc->sc_control_data,
912 	    sizeof(struct vge_control_data));
913  fail_2:
914 	bus_dmamem_free(sc->sc_dmat, &seg, nseg);
915  fail_1:
916 	return ENOMEM;
917 }
918 
919 /*
920  * Attach the interface. Allocate softc structures, do ifmedia
921  * setup and ethernet/BPF attach.
922  */
923 static void
924 vge_attach(struct device *parent, struct device *self, void *aux)
925 {
926 	uint8_t	*eaddr;
927 	struct vge_softc *sc = (void *)self;
928 	struct ifnet *ifp;
929 	struct pci_attach_args *pa = aux;
930 	pci_chipset_tag_t pc = pa->pa_pc;
931 	const char *intrstr;
932 	pci_intr_handle_t ih;
933 	uint16_t val;
934 
935 	aprint_normal(": VIA VT612X Gigabit Ethernet (rev. %#x)\n",
936 	    PCI_REVISION(pa->pa_class));
937 
938 	/* Make sure bus-mastering is enabled */
939         pci_conf_write(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG,
940 	    pci_conf_read(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG) |
941 	    PCI_COMMAND_MASTER_ENABLE);
942 
943 	/*
944 	 * Map control/status registers.
945 	 */
946 	if (pci_mapreg_map(pa, VGE_PCI_LOMEM, PCI_MAPREG_TYPE_MEM, 0,
947 	    &sc->sc_bst, &sc->sc_bsh, NULL, NULL) != 0) {
948 		aprint_error_dev(&sc->sc_dev, "couldn't map memory\n");
949 		return;
950 	}
951 
952         /*
953          * Map and establish our interrupt.
954          */
955 	if (pci_intr_map(pa, &ih)) {
956 		aprint_error_dev(&sc->sc_dev, "unable to map interrupt\n");
957 		return;
958 	}
959 	intrstr = pci_intr_string(pc, ih);
960 	sc->sc_intrhand = pci_intr_establish(pc, ih, IPL_NET, vge_intr, sc);
961 	if (sc->sc_intrhand == NULL) {
962 		aprint_error_dev(&sc->sc_dev, "unable to establish interrupt");
963 		if (intrstr != NULL)
964 			aprint_error(" at %s", intrstr);
965 		aprint_error("\n");
966 		return;
967 	}
968 	aprint_normal_dev(&sc->sc_dev, "interrupting at %s\n", intrstr);
969 
970 	/* Reset the adapter. */
971 	vge_reset(sc);
972 
973 	/*
974 	 * Get station address from the EEPROM.
975 	 */
976 	eaddr = sc->sc_eaddr;
977 	val = vge_read_eeprom(sc, VGE_EE_EADDR + 0);
978 	eaddr[0] = val & 0xff;
979 	eaddr[1] = val >> 8;
980 	val = vge_read_eeprom(sc, VGE_EE_EADDR + 1);
981 	eaddr[2] = val & 0xff;
982 	eaddr[3] = val >> 8;
983 	val = vge_read_eeprom(sc, VGE_EE_EADDR + 2);
984 	eaddr[4] = val & 0xff;
985 	eaddr[5] = val >> 8;
986 
987 	aprint_normal_dev(&sc->sc_dev, "Ethernet address: %s\n",
988 	    ether_sprintf(eaddr));
989 
990 	/*
991 	 * Use the 32bit tag. Hardware supports 48bit physical addresses,
992 	 * but we don't use that for now.
993 	 */
994 	sc->sc_dmat = pa->pa_dmat;
995 
996 	if (vge_allocmem(sc) != 0)
997 		return;
998 
999 	ifp = &sc->sc_ethercom.ec_if;
1000 	ifp->if_softc = sc;
1001 	strlcpy(ifp->if_xname, device_xname(&sc->sc_dev), IFNAMSIZ);
1002 	ifp->if_mtu = ETHERMTU;
1003 	ifp->if_baudrate = IF_Gbps(1);
1004 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
1005 	ifp->if_ioctl = vge_ioctl;
1006 	ifp->if_start = vge_start;
1007 	ifp->if_init = vge_init;
1008 	ifp->if_stop = vge_stop;
1009 
1010 	/*
1011 	 * We can support 802.1Q VLAN-sized frames and jumbo
1012 	 * Ethernet frames.
1013 	 */
1014 	sc->sc_ethercom.ec_capabilities |=
1015 	    ETHERCAP_VLAN_MTU | ETHERCAP_JUMBO_MTU |
1016 	    ETHERCAP_VLAN_HWTAGGING;
1017 
1018 	/*
1019 	 * We can do IPv4/TCPv4/UDPv4 checksums in hardware.
1020 	 */
1021 	ifp->if_capabilities |=
1022 	    IFCAP_CSUM_IPv4_Tx | IFCAP_CSUM_IPv4_Rx |
1023 	    IFCAP_CSUM_TCPv4_Tx | IFCAP_CSUM_TCPv4_Rx |
1024 	    IFCAP_CSUM_UDPv4_Tx | IFCAP_CSUM_UDPv4_Rx;
1025 
1026 #ifdef DEVICE_POLLING
1027 #ifdef IFCAP_POLLING
1028 	ifp->if_capabilities |= IFCAP_POLLING;
1029 #endif
1030 #endif
1031 	ifp->if_watchdog = vge_watchdog;
1032 	IFQ_SET_MAXLEN(&ifp->if_snd, max(VGE_IFQ_MAXLEN, IFQ_MAXLEN));
1033 	IFQ_SET_READY(&ifp->if_snd);
1034 
1035 	/*
1036 	 * Initialize our media structures and probe the MII.
1037 	 */
1038 	sc->sc_mii.mii_ifp = ifp;
1039 	sc->sc_mii.mii_readreg = vge_miibus_readreg;
1040 	sc->sc_mii.mii_writereg = vge_miibus_writereg;
1041 	sc->sc_mii.mii_statchg = vge_miibus_statchg;
1042 
1043 	sc->sc_ethercom.ec_mii = &sc->sc_mii;
1044 	ifmedia_init(&sc->sc_mii.mii_media, 0, ether_mediachange,
1045 	    ether_mediastatus);
1046 	mii_attach(&sc->sc_dev, &sc->sc_mii, 0xffffffff, MII_PHY_ANY,
1047 	    MII_OFFSET_ANY, MIIF_DOPAUSE);
1048 	if (LIST_FIRST(&sc->sc_mii.mii_phys) == NULL) {
1049 		ifmedia_add(&sc->sc_mii.mii_media, IFM_ETHER|IFM_NONE, 0, NULL);
1050 		ifmedia_set(&sc->sc_mii.mii_media, IFM_ETHER|IFM_NONE);
1051 	} else
1052 		ifmedia_set(&sc->sc_mii.mii_media, IFM_ETHER|IFM_AUTO);
1053 
1054 	/*
1055 	 * Attach the interface.
1056 	 */
1057 	if_attach(ifp);
1058 	ether_ifattach(ifp, eaddr);
1059 	ether_set_ifflags_cb(&sc->sc_ethercom, vge_ifflags_cb);
1060 
1061 	callout_init(&sc->sc_timeout, 0);
1062 	callout_setfunc(&sc->sc_timeout, vge_tick, sc);
1063 
1064 	/*
1065 	 * Make sure the interface is shutdown during reboot.
1066 	 */
1067 	if (shutdownhook_establish(vge_shutdown, sc) == NULL) {
1068 		aprint_error_dev(&sc->sc_dev, "WARNING: unable to establish shutdown hook\n");
1069 	}
1070 }
1071 
1072 static int
1073 vge_newbuf(struct vge_softc *sc, int idx, struct mbuf *m)
1074 {
1075 	struct mbuf *m_new;
1076 	struct vge_rxdesc *rxd;
1077 	struct vge_rxsoft *rxs;
1078 	bus_dmamap_t map;
1079 	int i;
1080 #ifdef DIAGNOSTIC
1081 	uint32_t rd_sts;
1082 #endif
1083 
1084 	m_new = NULL;
1085 	if (m == NULL) {
1086 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
1087 		if (m_new == NULL)
1088 			return ENOBUFS;
1089 
1090 		MCLGET(m_new, M_DONTWAIT);
1091 		if ((m_new->m_flags & M_EXT) == 0) {
1092 			m_freem(m_new);
1093 			return ENOBUFS;
1094 		}
1095 
1096 		m = m_new;
1097 	} else
1098 		m->m_data = m->m_ext.ext_buf;
1099 
1100 
1101 	/*
1102 	 * This is part of an evil trick to deal with non-x86 platforms.
1103 	 * The VIA chip requires RX buffers to be aligned on 32-bit
1104 	 * boundaries, but that will hose non-x86 machines. To get around
1105 	 * this, we leave some empty space at the start of each buffer
1106 	 * and for non-x86 hosts, we copy the buffer back two bytes
1107 	 * to achieve word alignment. This is slightly more efficient
1108 	 * than allocating a new buffer, copying the contents, and
1109 	 * discarding the old buffer.
1110 	 */
1111 	m->m_len = m->m_pkthdr.len = VGE_RX_BUFSIZE;
1112 #ifndef __NO_STRICT_ALIGNMENT
1113 	m->m_data += VGE_RX_PAD;
1114 #endif
1115 	rxs = &sc->sc_rxsoft[idx];
1116 	map = rxs->rxs_dmamap;
1117 
1118 	if (bus_dmamap_load_mbuf(sc->sc_dmat, map, m, BUS_DMA_NOWAIT) != 0)
1119 		goto out;
1120 
1121 	rxd = &sc->sc_rxdescs[idx];
1122 
1123 #ifdef DIAGNOSTIC
1124 	/* If this descriptor is still owned by the chip, bail. */
1125 	VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1126 	rd_sts = le32toh(rxd->rd_sts);
1127 	VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1128 	if (rd_sts & VGE_RDSTS_OWN) {
1129 		panic("%s: tried to map busy RX descriptor",
1130 		    device_xname(&sc->sc_dev));
1131 	}
1132 #endif
1133 
1134 	rxs->rxs_mbuf = m;
1135 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1136 	    BUS_DMASYNC_PREREAD);
1137 
1138 	rxd->rd_buflen =
1139 	    htole16(VGE_BUFLEN(map->dm_segs[0].ds_len) | VGE_RXDESC_I);
1140 	vge_set_rxaddr(rxd, map->dm_segs[0].ds_addr);
1141 	rxd->rd_sts = 0;
1142 	rxd->rd_ctl = 0;
1143 	VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1144 
1145 	/*
1146 	 * Note: the manual fails to document the fact that for
1147 	 * proper opration, the driver needs to replentish the RX
1148 	 * DMA ring 4 descriptors at a time (rather than one at a
1149 	 * time, like most chips). We can allocate the new buffers
1150 	 * but we should not set the OWN bits until we're ready
1151 	 * to hand back 4 of them in one shot.
1152 	 */
1153 
1154 #define VGE_RXCHUNK 4
1155 	sc->sc_rx_consumed++;
1156 	if (sc->sc_rx_consumed == VGE_RXCHUNK) {
1157 		for (i = idx; i != idx - VGE_RXCHUNK; i--) {
1158 			KASSERT(i >= 0);
1159 			sc->sc_rxdescs[i].rd_sts |= htole32(VGE_RDSTS_OWN);
1160 			VGE_RXDESCSYNC(sc, i,
1161 			    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1162 		}
1163 		sc->sc_rx_consumed = 0;
1164 	}
1165 
1166 	return 0;
1167  out:
1168 	if (m_new != NULL)
1169 		m_freem(m_new);
1170 	return ENOMEM;
1171 }
1172 
1173 #ifndef __NO_STRICT_ALIGNMENT
1174 static inline void
1175 vge_fixup_rx(struct mbuf *m)
1176 {
1177 	int i;
1178 	uint16_t *src, *dst;
1179 
1180 	src = mtod(m, uint16_t *);
1181 	dst = src - 1;
1182 
1183 	for (i = 0; i < (m->m_len / sizeof(uint16_t) + 1); i++)
1184 		*dst++ = *src++;
1185 
1186 	m->m_data -= ETHER_ALIGN;
1187 }
1188 #endif
1189 
1190 /*
1191  * RX handler. We support the reception of jumbo frames that have
1192  * been fragmented across multiple 2K mbuf cluster buffers.
1193  */
1194 static void
1195 vge_rxeof(struct vge_softc *sc)
1196 {
1197 	struct mbuf *m;
1198 	struct ifnet *ifp;
1199 	int idx, total_len, lim;
1200 	struct vge_rxdesc *cur_rxd;
1201 	struct vge_rxsoft *rxs;
1202 	uint32_t rxstat, rxctl;
1203 
1204 	ifp = &sc->sc_ethercom.ec_if;
1205 	lim = 0;
1206 
1207 	/* Invalidate the descriptor memory */
1208 
1209 	for (idx = sc->sc_rx_prodidx;; idx = VGE_NEXT_RXDESC(idx)) {
1210 		cur_rxd = &sc->sc_rxdescs[idx];
1211 
1212 		VGE_RXDESCSYNC(sc, idx,
1213 		    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1214 		rxstat = le32toh(cur_rxd->rd_sts);
1215 		if ((rxstat & VGE_RDSTS_OWN) != 0) {
1216 			VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1217 			break;
1218 		}
1219 
1220 		rxctl = le32toh(cur_rxd->rd_ctl);
1221 		rxs = &sc->sc_rxsoft[idx];
1222 		m = rxs->rxs_mbuf;
1223 		total_len = (rxstat & VGE_RDSTS_BUFSIZ) >> 16;
1224 
1225 		/* Invalidate the RX mbuf and unload its map */
1226 
1227 		bus_dmamap_sync(sc->sc_dmat, rxs->rxs_dmamap,
1228 		    0, rxs->rxs_dmamap->dm_mapsize, BUS_DMASYNC_POSTREAD);
1229 		bus_dmamap_unload(sc->sc_dmat, rxs->rxs_dmamap);
1230 
1231 		/*
1232 		 * If the 'start of frame' bit is set, this indicates
1233 		 * either the first fragment in a multi-fragment receive,
1234 		 * or an intermediate fragment. Either way, we want to
1235 		 * accumulate the buffers.
1236 		 */
1237 		if (rxstat & VGE_RXPKT_SOF) {
1238 			m->m_len = VGE_RX_BUFSIZE;
1239 			if (sc->sc_rx_mhead == NULL)
1240 				sc->sc_rx_mhead = sc->sc_rx_mtail = m;
1241 			else {
1242 				m->m_flags &= ~M_PKTHDR;
1243 				sc->sc_rx_mtail->m_next = m;
1244 				sc->sc_rx_mtail = m;
1245 			}
1246 			vge_newbuf(sc, idx, NULL);
1247 			continue;
1248 		}
1249 
1250 		/*
1251 		 * Bad/error frames will have the RXOK bit cleared.
1252 		 * However, there's one error case we want to allow:
1253 		 * if a VLAN tagged frame arrives and the chip can't
1254 		 * match it against the CAM filter, it considers this
1255 		 * a 'VLAN CAM filter miss' and clears the 'RXOK' bit.
1256 		 * We don't want to drop the frame though: our VLAN
1257 		 * filtering is done in software.
1258 		 */
1259 		if ((rxstat & VGE_RDSTS_RXOK) == 0 &&
1260 		    (rxstat & VGE_RDSTS_VIDM) == 0 &&
1261 		    (rxstat & VGE_RDSTS_CSUMERR) == 0) {
1262 			ifp->if_ierrors++;
1263 			/*
1264 			 * If this is part of a multi-fragment packet,
1265 			 * discard all the pieces.
1266 			 */
1267 			if (sc->sc_rx_mhead != NULL) {
1268 				m_freem(sc->sc_rx_mhead);
1269 				sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1270 			}
1271 			vge_newbuf(sc, idx, m);
1272 			continue;
1273 		}
1274 
1275 		/*
1276 		 * If allocating a replacement mbuf fails,
1277 		 * reload the current one.
1278 		 */
1279 
1280 		if (vge_newbuf(sc, idx, NULL)) {
1281 			ifp->if_ierrors++;
1282 			if (sc->sc_rx_mhead != NULL) {
1283 				m_freem(sc->sc_rx_mhead);
1284 				sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1285 			}
1286 			vge_newbuf(sc, idx, m);
1287 			continue;
1288 		}
1289 
1290 		if (sc->sc_rx_mhead != NULL) {
1291 			m->m_len = total_len % VGE_RX_BUFSIZE;
1292 			/*
1293 			 * Special case: if there's 4 bytes or less
1294 			 * in this buffer, the mbuf can be discarded:
1295 			 * the last 4 bytes is the CRC, which we don't
1296 			 * care about anyway.
1297 			 */
1298 			if (m->m_len <= ETHER_CRC_LEN) {
1299 				sc->sc_rx_mtail->m_len -=
1300 				    (ETHER_CRC_LEN - m->m_len);
1301 				m_freem(m);
1302 			} else {
1303 				m->m_len -= ETHER_CRC_LEN;
1304 				m->m_flags &= ~M_PKTHDR;
1305 				sc->sc_rx_mtail->m_next = m;
1306 			}
1307 			m = sc->sc_rx_mhead;
1308 			sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1309 			m->m_pkthdr.len = total_len - ETHER_CRC_LEN;
1310 		} else
1311 			m->m_pkthdr.len = m->m_len = total_len - ETHER_CRC_LEN;
1312 
1313 #ifndef __NO_STRICT_ALIGNMENT
1314 		vge_fixup_rx(m);
1315 #endif
1316 		ifp->if_ipackets++;
1317 		m->m_pkthdr.rcvif = ifp;
1318 
1319 		/* Do RX checksumming if enabled */
1320 		if (ifp->if_csum_flags_rx & M_CSUM_IPv4) {
1321 
1322 			/* Check IP header checksum */
1323 			if (rxctl & VGE_RDCTL_IPPKT)
1324 				m->m_pkthdr.csum_flags |= M_CSUM_IPv4;
1325 			if ((rxctl & VGE_RDCTL_IPCSUMOK) == 0)
1326 				m->m_pkthdr.csum_flags |= M_CSUM_IPv4_BAD;
1327 		}
1328 
1329 		if (ifp->if_csum_flags_rx & M_CSUM_TCPv4) {
1330 			/* Check UDP checksum */
1331 			if (rxctl & VGE_RDCTL_TCPPKT)
1332 				m->m_pkthdr.csum_flags |= M_CSUM_TCPv4;
1333 
1334 			if ((rxctl & VGE_RDCTL_PROTOCSUMOK) == 0)
1335 				m->m_pkthdr.csum_flags |= M_CSUM_TCP_UDP_BAD;
1336 		}
1337 
1338 		if (ifp->if_csum_flags_rx & M_CSUM_UDPv4) {
1339 			/* Check UDP checksum */
1340 			if (rxctl & VGE_RDCTL_UDPPKT)
1341 				m->m_pkthdr.csum_flags |= M_CSUM_UDPv4;
1342 
1343 			if ((rxctl & VGE_RDCTL_PROTOCSUMOK) == 0)
1344 				m->m_pkthdr.csum_flags |= M_CSUM_TCP_UDP_BAD;
1345 		}
1346 
1347 		if (rxstat & VGE_RDSTS_VTAG) {
1348 			/*
1349 			 * We use bswap16() here because:
1350 			 * On LE machines, tag is stored in BE as stream data.
1351 			 * On BE machines, tag is stored in BE as stream data
1352 			 *  but it was already swapped by le32toh() above.
1353 			 */
1354 			VLAN_INPUT_TAG(ifp, m,
1355 			    bswap16(rxctl & VGE_RDCTL_VLANID), continue);
1356 		}
1357 
1358 #if NBPFILTER > 0
1359 		/*
1360 		 * Handle BPF listeners.
1361 		 */
1362 		if (ifp->if_bpf)
1363 			bpf_mtap(ifp->if_bpf, m);
1364 #endif
1365 
1366 		(*ifp->if_input)(ifp, m);
1367 
1368 		lim++;
1369 		if (lim == VGE_NRXDESC)
1370 			break;
1371 	}
1372 
1373 	sc->sc_rx_prodidx = idx;
1374 	CSR_WRITE_2(sc, VGE_RXDESC_RESIDUECNT, lim);
1375 }
1376 
1377 static void
1378 vge_txeof(struct vge_softc *sc)
1379 {
1380 	struct ifnet *ifp;
1381 	struct vge_txsoft *txs;
1382 	uint32_t txstat;
1383 	int idx;
1384 
1385 	ifp = &sc->sc_ethercom.ec_if;
1386 
1387 	for (idx = sc->sc_tx_considx;
1388 	    sc->sc_tx_free < VGE_NTXDESC;
1389 	    idx = VGE_NEXT_TXDESC(idx), sc->sc_tx_free++) {
1390 		VGE_TXDESCSYNC(sc, idx,
1391 		    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1392 		txstat = le32toh(sc->sc_txdescs[idx].td_sts);
1393 		VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1394 		if (txstat & VGE_TDSTS_OWN) {
1395 			break;
1396 		}
1397 
1398 		txs = &sc->sc_txsoft[idx];
1399 		m_freem(txs->txs_mbuf);
1400 		txs->txs_mbuf = NULL;
1401 		bus_dmamap_sync(sc->sc_dmat, txs->txs_dmamap, 0,
1402 		    txs->txs_dmamap->dm_mapsize, BUS_DMASYNC_POSTWRITE);
1403 		bus_dmamap_unload(sc->sc_dmat, txs->txs_dmamap);
1404 		if (txstat & (VGE_TDSTS_EXCESSCOLL|VGE_TDSTS_COLL))
1405 			ifp->if_collisions++;
1406 		if (txstat & VGE_TDSTS_TXERR)
1407 			ifp->if_oerrors++;
1408 		else
1409 			ifp->if_opackets++;
1410 	}
1411 
1412 	sc->sc_tx_considx = idx;
1413 
1414 	if (sc->sc_tx_free > 0) {
1415 		ifp->if_flags &= ~IFF_OACTIVE;
1416 	}
1417 
1418 	/*
1419 	 * If not all descriptors have been released reaped yet,
1420 	 * reload the timer so that we will eventually get another
1421 	 * interrupt that will cause us to re-enter this routine.
1422 	 * This is done in case the transmitter has gone idle.
1423 	 */
1424 	if (sc->sc_tx_free < VGE_NTXDESC)
1425 		CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_TIMER0_ENABLE);
1426 	else
1427 		ifp->if_timer = 0;
1428 }
1429 
1430 static void
1431 vge_tick(void *xsc)
1432 {
1433 	struct vge_softc *sc;
1434 	struct ifnet *ifp;
1435 	struct mii_data *mii;
1436 	int s;
1437 
1438 	sc = xsc;
1439 	ifp = &sc->sc_ethercom.ec_if;
1440 	mii = &sc->sc_mii;
1441 
1442 	s = splnet();
1443 
1444 	callout_schedule(&sc->sc_timeout, hz);
1445 
1446 	mii_tick(mii);
1447 	if (sc->sc_link) {
1448 		if ((mii->mii_media_status & IFM_ACTIVE) == 0)
1449 			sc->sc_link = 0;
1450 	} else {
1451 		if (mii->mii_media_status & IFM_ACTIVE &&
1452 		    IFM_SUBTYPE(mii->mii_media_active) != IFM_NONE) {
1453 			sc->sc_link = 1;
1454 			if (!IFQ_IS_EMPTY(&ifp->if_snd))
1455 				vge_start(ifp);
1456 		}
1457 	}
1458 
1459 	splx(s);
1460 }
1461 
1462 static int
1463 vge_intr(void *arg)
1464 {
1465 	struct vge_softc *sc;
1466 	struct ifnet *ifp;
1467 	uint32_t status;
1468 	int claim;
1469 
1470 	sc = arg;
1471 	claim = 0;
1472 	if (sc->sc_suspended) {
1473 		return claim;
1474 	}
1475 
1476 	ifp = &sc->sc_ethercom.ec_if;
1477 
1478 	if ((ifp->if_flags & IFF_UP) == 0) {
1479 		return claim;
1480 	}
1481 
1482 	/* Disable interrupts */
1483 	CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
1484 
1485 	for (;;) {
1486 
1487 		status = CSR_READ_4(sc, VGE_ISR);
1488 		/* If the card has gone away the read returns 0xffff. */
1489 		if (status == 0xFFFFFFFF)
1490 			break;
1491 
1492 		if (status) {
1493 			claim = 1;
1494 			CSR_WRITE_4(sc, VGE_ISR, status);
1495 		}
1496 
1497 		if ((status & VGE_INTRS) == 0)
1498 			break;
1499 
1500 		if (status & (VGE_ISR_RXOK|VGE_ISR_RXOK_HIPRIO))
1501 			vge_rxeof(sc);
1502 
1503 		if (status & (VGE_ISR_RXOFLOW|VGE_ISR_RXNODESC)) {
1504 			vge_rxeof(sc);
1505 			CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_RUN);
1506 			CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_WAK);
1507 		}
1508 
1509 		if (status & (VGE_ISR_TXOK0|VGE_ISR_TIMER0))
1510 			vge_txeof(sc);
1511 
1512 		if (status & (VGE_ISR_TXDMA_STALL|VGE_ISR_RXDMA_STALL))
1513 			vge_init(ifp);
1514 
1515 		if (status & VGE_ISR_LINKSTS)
1516 			vge_tick(sc);
1517 	}
1518 
1519 	/* Re-enable interrupts */
1520 	CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_GMSK);
1521 
1522 	if (claim && !IFQ_IS_EMPTY(&ifp->if_snd))
1523 		vge_start(ifp);
1524 
1525 	return claim;
1526 }
1527 
1528 static int
1529 vge_encap(struct vge_softc *sc, struct mbuf *m_head, int idx)
1530 {
1531 	struct vge_txsoft *txs;
1532 	struct vge_txdesc *txd;
1533 	struct vge_txfrag *f;
1534 	struct mbuf *m_new;
1535 	bus_dmamap_t map;
1536 	int m_csumflags, seg, error, flags;
1537 	struct m_tag *mtag;
1538 	size_t sz;
1539 	uint32_t td_sts, td_ctl;
1540 
1541 	KASSERT(sc->sc_tx_free > 0);
1542 
1543 	txd = &sc->sc_txdescs[idx];
1544 
1545 #ifdef DIAGNOSTIC
1546 	/* If this descriptor is still owned by the chip, bail. */
1547 	VGE_TXDESCSYNC(sc, idx,
1548 	    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1549 	td_sts = le32toh(txd->td_sts);
1550 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1551 	if (td_sts & VGE_TDSTS_OWN) {
1552 		return ENOBUFS;
1553 	}
1554 #endif
1555 
1556 	/*
1557 	 * Preserve m_pkthdr.csum_flags here since m_head might be
1558 	 * updated by m_defrag()
1559 	 */
1560 	m_csumflags = m_head->m_pkthdr.csum_flags;
1561 
1562 	txs = &sc->sc_txsoft[idx];
1563 	map = txs->txs_dmamap;
1564 	error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m_head, BUS_DMA_NOWAIT);
1565 
1566 	/* If too many segments to map, coalesce */
1567 	if (error == EFBIG ||
1568 	    (m_head->m_pkthdr.len < ETHER_PAD_LEN &&
1569 	     map->dm_nsegs == VGE_TX_FRAGS)) {
1570 		m_new = m_defrag(m_head, M_DONTWAIT);
1571 		if (m_new == NULL)
1572 			return EFBIG;
1573 
1574 		error = bus_dmamap_load_mbuf(sc->sc_dmat, map,
1575 		    m_new, BUS_DMA_NOWAIT);
1576 		if (error) {
1577 			m_freem(m_new);
1578 			return error;
1579 		}
1580 
1581 		m_head = m_new;
1582 	} else if (error)
1583 		return error;
1584 
1585 	txs->txs_mbuf = m_head;
1586 
1587 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1588 	    BUS_DMASYNC_PREWRITE);
1589 
1590 	for (seg = 0, f = &txd->td_frag[0]; seg < map->dm_nsegs; seg++, f++) {
1591 		f->tf_buflen = htole16(VGE_BUFLEN(map->dm_segs[seg].ds_len));
1592 		vge_set_txaddr(f, map->dm_segs[seg].ds_addr);
1593 	}
1594 
1595 	/* Argh. This chip does not autopad short frames */
1596 	sz = m_head->m_pkthdr.len;
1597 	if (sz < ETHER_PAD_LEN) {
1598 		f->tf_buflen = htole16(VGE_BUFLEN(ETHER_PAD_LEN - sz));
1599 		vge_set_txaddr(f, VGE_CDPADADDR(sc));
1600 		sz = ETHER_PAD_LEN;
1601 		seg++;
1602 	}
1603 	VGE_TXFRAGSYNC(sc, idx, seg, BUS_DMASYNC_PREWRITE);
1604 
1605 	/*
1606 	 * When telling the chip how many segments there are, we
1607 	 * must use nsegs + 1 instead of just nsegs. Darned if I
1608 	 * know why.
1609 	 */
1610 	seg++;
1611 
1612 	flags = 0;
1613 	if (m_csumflags & M_CSUM_IPv4)
1614 		flags |= VGE_TDCTL_IPCSUM;
1615 	if (m_csumflags & M_CSUM_TCPv4)
1616 		flags |= VGE_TDCTL_TCPCSUM;
1617 	if (m_csumflags & M_CSUM_UDPv4)
1618 		flags |= VGE_TDCTL_UDPCSUM;
1619 	td_sts = sz << 16;
1620 	td_ctl = flags | (seg << 28) | VGE_TD_LS_NORM;
1621 
1622 	if (sz > ETHERMTU + ETHER_HDR_LEN)
1623 		td_ctl |= VGE_TDCTL_JUMBO;
1624 
1625 	/*
1626 	 * Set up hardware VLAN tagging.
1627 	 */
1628 	mtag = VLAN_OUTPUT_TAG(&sc->sc_ethercom, m_head);
1629 	if (mtag != NULL) {
1630 		/*
1631 		 * No need htons() here since vge(4) chip assumes
1632 		 * that tags are written in little endian and
1633 		 * we already use htole32() here.
1634 		 */
1635 		td_ctl |= VLAN_TAG_VALUE(mtag) | VGE_TDCTL_VTAG;
1636 	}
1637 	txd->td_ctl = htole32(td_ctl);
1638 	txd->td_sts = htole32(td_sts);
1639 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1640 
1641 	txd->td_sts = htole32(VGE_TDSTS_OWN | td_sts);
1642 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1643 
1644 	sc->sc_tx_free--;
1645 
1646 	return 0;
1647 }
1648 
1649 /*
1650  * Main transmit routine.
1651  */
1652 
1653 static void
1654 vge_start(struct ifnet *ifp)
1655 {
1656 	struct vge_softc *sc;
1657 	struct vge_txsoft *txs;
1658 	struct mbuf *m_head;
1659 	int idx, pidx, ofree, error;
1660 
1661 	sc = ifp->if_softc;
1662 
1663 	if (!sc->sc_link ||
1664 	    (ifp->if_flags & (IFF_RUNNING|IFF_OACTIVE)) != IFF_RUNNING) {
1665 		return;
1666 	}
1667 
1668 	m_head = NULL;
1669 	idx = sc->sc_tx_prodidx;
1670 	pidx = VGE_PREV_TXDESC(idx);
1671 	ofree = sc->sc_tx_free;
1672 
1673 	/*
1674 	 * Loop through the send queue, setting up transmit descriptors
1675 	 * until we drain the queue, or use up all available transmit
1676 	 * descriptors.
1677 	 */
1678 	for (;;) {
1679 		/* Grab a packet off the queue. */
1680 		IFQ_POLL(&ifp->if_snd, m_head);
1681 		if (m_head == NULL)
1682 			break;
1683 
1684 		if (sc->sc_tx_free == 0) {
1685 			/*
1686 			 * All slots used, stop for now.
1687 			 */
1688 			ifp->if_flags |= IFF_OACTIVE;
1689 			break;
1690 		}
1691 
1692 		txs = &sc->sc_txsoft[idx];
1693 		KASSERT(txs->txs_mbuf == NULL);
1694 
1695 		if ((error = vge_encap(sc, m_head, idx))) {
1696 			if (error == EFBIG) {
1697 				aprint_error_dev(&sc->sc_dev, "Tx packet consumes too many "
1698 				    "DMA segments, dropping...\n");
1699 				IFQ_DEQUEUE(&ifp->if_snd, m_head);
1700 				m_freem(m_head);
1701 				continue;
1702 			}
1703 
1704 			/*
1705 			 * Short on resources, just stop for now.
1706 			 */
1707 			if (error == ENOBUFS)
1708 				ifp->if_flags |= IFF_OACTIVE;
1709 			break;
1710 		}
1711 
1712 		IFQ_DEQUEUE(&ifp->if_snd, m_head);
1713 
1714 		/*
1715 		 * WE ARE NOW COMMITTED TO TRANSMITTING THE PACKET.
1716 		 */
1717 
1718 		sc->sc_txdescs[pidx].td_frag[0].tf_buflen |=
1719 		    htole16(VGE_TXDESC_Q);
1720 		VGE_TXFRAGSYNC(sc, pidx, 1,
1721 		    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1722 
1723 		if (txs->txs_mbuf != m_head) {
1724 			m_freem(m_head);
1725 			m_head = txs->txs_mbuf;
1726 		}
1727 
1728 		pidx = idx;
1729 		idx = VGE_NEXT_TXDESC(idx);
1730 
1731 		/*
1732 		 * If there's a BPF listener, bounce a copy of this frame
1733 		 * to him.
1734 		 */
1735 #if NBPFILTER > 0
1736 		if (ifp->if_bpf)
1737 			bpf_mtap(ifp->if_bpf, m_head);
1738 #endif
1739 	}
1740 
1741 	if (sc->sc_tx_free < ofree) {
1742 		/* TX packet queued */
1743 
1744 		sc->sc_tx_prodidx = idx;
1745 
1746 		/* Issue a transmit command. */
1747 		CSR_WRITE_2(sc, VGE_TXQCSRS, VGE_TXQCSR_WAK0);
1748 
1749 		/*
1750 		 * Use the countdown timer for interrupt moderation.
1751 		 * 'TX done' interrupts are disabled. Instead, we reset the
1752 		 * countdown timer, which will begin counting until it hits
1753 		 * the value in the SSTIMER register, and then trigger an
1754 		 * interrupt. Each time we set the TIMER0_ENABLE bit, the
1755 		 * the timer count is reloaded. Only when the transmitter
1756 		 * is idle will the timer hit 0 and an interrupt fire.
1757 		 */
1758 		CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_TIMER0_ENABLE);
1759 
1760 		/*
1761 		 * Set a timeout in case the chip goes out to lunch.
1762 		 */
1763 		ifp->if_timer = 5;
1764 	}
1765 }
1766 
1767 static int
1768 vge_init(struct ifnet *ifp)
1769 {
1770 	struct vge_softc *sc;
1771 	int i, rc = 0;
1772 
1773 	sc = ifp->if_softc;
1774 
1775 	/*
1776 	 * Cancel pending I/O and free all RX/TX buffers.
1777 	 */
1778 	vge_stop(ifp, 0);
1779 	vge_reset(sc);
1780 
1781 	/* Initialize the RX descriptors and mbufs. */
1782 	memset(sc->sc_rxdescs, 0, sizeof(sc->sc_rxdescs));
1783 	sc->sc_rx_consumed = 0;
1784 	for (i = 0; i < VGE_NRXDESC; i++) {
1785 		if (vge_newbuf(sc, i, NULL) == ENOBUFS) {
1786 			aprint_error_dev(&sc->sc_dev, "unable to allocate or map "
1787 			    "rx buffer\n");
1788 			return 1; /* XXX */
1789 		}
1790 	}
1791 	sc->sc_rx_prodidx = 0;
1792 	sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1793 
1794 	/* Initialize the  TX descriptors and mbufs. */
1795 	memset(sc->sc_txdescs, 0, sizeof(sc->sc_txdescs));
1796 	bus_dmamap_sync(sc->sc_dmat, sc->sc_cddmamap,
1797 	    VGE_CDTXOFF(0), sizeof(sc->sc_txdescs),
1798 	    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1799 	for (i = 0; i < VGE_NTXDESC; i++)
1800 		sc->sc_txsoft[i].txs_mbuf = NULL;
1801 
1802 	sc->sc_tx_prodidx = 0;
1803 	sc->sc_tx_considx = 0;
1804 	sc->sc_tx_free = VGE_NTXDESC;
1805 
1806 	/* Set our station address */
1807 	for (i = 0; i < ETHER_ADDR_LEN; i++)
1808 		CSR_WRITE_1(sc, VGE_PAR0 + i, sc->sc_eaddr[i]);
1809 
1810 	/*
1811 	 * Set receive FIFO threshold. Also allow transmission and
1812 	 * reception of VLAN tagged frames.
1813 	 */
1814 	CSR_CLRBIT_1(sc, VGE_RXCFG, VGE_RXCFG_FIFO_THR|VGE_RXCFG_VTAGOPT);
1815 	CSR_SETBIT_1(sc, VGE_RXCFG, VGE_RXFIFOTHR_128BYTES|VGE_VTAG_OPT2);
1816 
1817 	/* Set DMA burst length */
1818 	CSR_CLRBIT_1(sc, VGE_DMACFG0, VGE_DMACFG0_BURSTLEN);
1819 	CSR_SETBIT_1(sc, VGE_DMACFG0, VGE_DMABURST_128);
1820 
1821 	CSR_SETBIT_1(sc, VGE_TXCFG, VGE_TXCFG_ARB_PRIO|VGE_TXCFG_NONBLK);
1822 
1823 	/* Set collision backoff algorithm */
1824 	CSR_CLRBIT_1(sc, VGE_CHIPCFG1, VGE_CHIPCFG1_CRANDOM|
1825 	    VGE_CHIPCFG1_CAP|VGE_CHIPCFG1_MBA|VGE_CHIPCFG1_BAKOPT);
1826 	CSR_SETBIT_1(sc, VGE_CHIPCFG1, VGE_CHIPCFG1_OFSET);
1827 
1828 	/* Disable LPSEL field in priority resolution */
1829 	CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_LPSEL_DIS);
1830 
1831 	/*
1832 	 * Load the addresses of the DMA queues into the chip.
1833 	 * Note that we only use one transmit queue.
1834 	 */
1835 
1836 	CSR_WRITE_4(sc, VGE_TXDESC_ADDR_LO0, VGE_ADDR_LO(VGE_CDTXADDR(sc, 0)));
1837 	CSR_WRITE_2(sc, VGE_TXDESCNUM, VGE_NTXDESC - 1);
1838 
1839 	CSR_WRITE_4(sc, VGE_RXDESC_ADDR_LO, VGE_ADDR_LO(VGE_CDRXADDR(sc, 0)));
1840 	CSR_WRITE_2(sc, VGE_RXDESCNUM, VGE_NRXDESC - 1);
1841 	CSR_WRITE_2(sc, VGE_RXDESC_RESIDUECNT, VGE_NRXDESC);
1842 
1843 	/* Enable and wake up the RX descriptor queue */
1844 	CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_RUN);
1845 	CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_WAK);
1846 
1847 	/* Enable the TX descriptor queue */
1848 	CSR_WRITE_2(sc, VGE_TXQCSRS, VGE_TXQCSR_RUN0);
1849 
1850 	/* Set up the receive filter -- allow large frames for VLANs. */
1851 	CSR_WRITE_1(sc, VGE_RXCTL, VGE_RXCTL_RX_UCAST|VGE_RXCTL_RX_GIANT);
1852 
1853 	/* If we want promiscuous mode, set the allframes bit. */
1854 	if (ifp->if_flags & IFF_PROMISC) {
1855 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_PROMISC);
1856 	}
1857 
1858 	/* Set capture broadcast bit to capture broadcast frames. */
1859 	if (ifp->if_flags & IFF_BROADCAST) {
1860 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_BCAST);
1861 	}
1862 
1863 	/* Set multicast bit to capture multicast frames. */
1864 	if (ifp->if_flags & IFF_MULTICAST) {
1865 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_MCAST);
1866 	}
1867 
1868 	/* Init the cam filter. */
1869 	vge_cam_clear(sc);
1870 
1871 	/* Init the multicast filter. */
1872 	vge_setmulti(sc);
1873 
1874 	/* Enable flow control */
1875 
1876 	CSR_WRITE_1(sc, VGE_CRS2, 0x8B);
1877 
1878 	/* Enable jumbo frame reception (if desired) */
1879 
1880 	/* Start the MAC. */
1881 	CSR_WRITE_1(sc, VGE_CRC0, VGE_CR0_STOP);
1882 	CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_NOPOLL);
1883 	CSR_WRITE_1(sc, VGE_CRS0,
1884 	    VGE_CR0_TX_ENABLE|VGE_CR0_RX_ENABLE|VGE_CR0_START);
1885 
1886 	/*
1887 	 * Configure one-shot timer for microsecond
1888 	 * resulution and load it for 500 usecs.
1889 	 */
1890 	CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_TIMER0_RES);
1891 	CSR_WRITE_2(sc, VGE_SSTIMER, 400);
1892 
1893 	/*
1894 	 * Configure interrupt moderation for receive. Enable
1895 	 * the holdoff counter and load it, and set the RX
1896 	 * suppression count to the number of descriptors we
1897 	 * want to allow before triggering an interrupt.
1898 	 * The holdoff timer is in units of 20 usecs.
1899 	 */
1900 
1901 #ifdef notyet
1902 	CSR_WRITE_1(sc, VGE_INTCTL1, VGE_INTCTL_TXINTSUP_DISABLE);
1903 	/* Select the interrupt holdoff timer page. */
1904 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1905 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_INTHLDOFF);
1906 	CSR_WRITE_1(sc, VGE_INTHOLDOFF, 10); /* ~200 usecs */
1907 
1908 	/* Enable use of the holdoff timer. */
1909 	CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_HOLDOFF);
1910 	CSR_WRITE_1(sc, VGE_INTCTL1, VGE_INTCTL_SC_RELOAD);
1911 
1912 	/* Select the RX suppression threshold page. */
1913 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1914 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_RXSUPPTHR);
1915 	CSR_WRITE_1(sc, VGE_RXSUPPTHR, 64); /* interrupt after 64 packets */
1916 
1917 	/* Restore the page select bits. */
1918 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1919 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_MAR);
1920 #endif
1921 
1922 #ifdef DEVICE_POLLING
1923 	/*
1924 	 * Disable interrupts if we are polling.
1925 	 */
1926 	if (ifp->if_flags & IFF_POLLING) {
1927 		CSR_WRITE_4(sc, VGE_IMR, 0);
1928 		CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
1929 	} else	/* otherwise ... */
1930 #endif /* DEVICE_POLLING */
1931 	{
1932 	/*
1933 	 * Enable interrupts.
1934 	 */
1935 		CSR_WRITE_4(sc, VGE_IMR, VGE_INTRS);
1936 		CSR_WRITE_4(sc, VGE_ISR, 0);
1937 		CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_GMSK);
1938 	}
1939 
1940 	if ((rc = ether_mediachange(ifp)) != 0)
1941 		goto out;
1942 
1943 	ifp->if_flags |= IFF_RUNNING;
1944 	ifp->if_flags &= ~IFF_OACTIVE;
1945 
1946 	sc->sc_if_flags = 0;
1947 	sc->sc_link = 0;
1948 
1949 	callout_schedule(&sc->sc_timeout, hz);
1950 
1951 out:
1952 	return rc;
1953 }
1954 
1955 static void
1956 vge_miibus_statchg(struct device *self)
1957 {
1958 	struct vge_softc *sc;
1959 	struct mii_data *mii;
1960 	struct ifmedia_entry *ife;
1961 
1962 	sc = (void *)self;
1963 	mii = &sc->sc_mii;
1964 	ife = mii->mii_media.ifm_cur;
1965 	/*
1966 	 * If the user manually selects a media mode, we need to turn
1967 	 * on the forced MAC mode bit in the DIAGCTL register. If the
1968 	 * user happens to choose a full duplex mode, we also need to
1969 	 * set the 'force full duplex' bit. This applies only to
1970 	 * 10Mbps and 100Mbps speeds. In autoselect mode, forced MAC
1971 	 * mode is disabled, and in 1000baseT mode, full duplex is
1972 	 * always implied, so we turn on the forced mode bit but leave
1973 	 * the FDX bit cleared.
1974 	 */
1975 
1976 	switch (IFM_SUBTYPE(ife->ifm_media)) {
1977 	case IFM_AUTO:
1978 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
1979 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1980 		break;
1981 	case IFM_1000_T:
1982 		CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
1983 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1984 		break;
1985 	case IFM_100_TX:
1986 	case IFM_10_T:
1987 		CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
1988 		if ((ife->ifm_media & IFM_GMASK) == IFM_FDX) {
1989 			CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1990 		} else {
1991 			CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1992 		}
1993 		break;
1994 	default:
1995 		aprint_error_dev(&sc->sc_dev, "unknown media type: %x\n",
1996 		    IFM_SUBTYPE(ife->ifm_media));
1997 		break;
1998 	}
1999 }
2000 
2001 static int
2002 vge_ifflags_cb(struct ethercom *ec)
2003 {
2004 	struct ifnet *ifp = &ec->ec_if;
2005 	struct vge_softc *sc = ifp->if_softc;
2006 	int change = ifp->if_flags ^ sc->sc_if_flags;
2007 
2008 	if ((change & ~(IFF_CANTCHANGE|IFF_DEBUG)) != 0)
2009 		return ENETRESET;
2010 	else if ((change & IFF_PROMISC) == 0)
2011 		return 0;
2012 
2013 	if ((ifp->if_flags & IFF_PROMISC) == 0)
2014 		CSR_CLRBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_PROMISC);
2015 	else
2016 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_PROMISC);
2017 	vge_setmulti(sc);
2018 	return 0;
2019 }
2020 
2021 static int
2022 vge_ioctl(struct ifnet *ifp, u_long command, void *data)
2023 {
2024 	struct vge_softc *sc;
2025 	struct ifreq *ifr;
2026 	int s, error;
2027 
2028 	sc = ifp->if_softc;
2029 	ifr = (struct ifreq *)data;
2030 	error = 0;
2031 
2032 	s = splnet();
2033 
2034 	if ((error = ether_ioctl(ifp, command, data)) == ENETRESET) {
2035 		error = 0;
2036 		if (command != SIOCADDMULTI && command != SIOCDELMULTI)
2037 			;
2038 		else if (ifp->if_flags & IFF_RUNNING) {
2039 			/*
2040 			 * Multicast list has changed; set the hardware filter
2041 			 * accordingly.
2042 			 */
2043 			vge_setmulti(sc);
2044 		}
2045 	}
2046 	sc->sc_if_flags = ifp->if_flags;
2047 
2048 	splx(s);
2049 	return error;
2050 }
2051 
2052 static void
2053 vge_watchdog(struct ifnet *ifp)
2054 {
2055 	struct vge_softc *sc;
2056 	int s;
2057 
2058 	sc = ifp->if_softc;
2059 	s = splnet();
2060 	aprint_error_dev(&sc->sc_dev, "watchdog timeout\n");
2061 	ifp->if_oerrors++;
2062 
2063 	vge_txeof(sc);
2064 	vge_rxeof(sc);
2065 
2066 	vge_init(ifp);
2067 
2068 	splx(s);
2069 }
2070 
2071 /*
2072  * Stop the adapter and free any mbufs allocated to the
2073  * RX and TX lists.
2074  */
2075 static void
2076 vge_stop(struct ifnet *ifp, int disable)
2077 {
2078 	struct vge_softc *sc = ifp->if_softc;
2079 	struct vge_txsoft *txs;
2080 	struct vge_rxsoft *rxs;
2081 	int i, s;
2082 
2083 	s = splnet();
2084 	ifp->if_timer = 0;
2085 
2086 	ifp->if_flags &= ~(IFF_RUNNING | IFF_OACTIVE);
2087 #ifdef DEVICE_POLLING
2088 	ether_poll_deregister(ifp);
2089 #endif /* DEVICE_POLLING */
2090 
2091 	CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
2092 	CSR_WRITE_1(sc, VGE_CRS0, VGE_CR0_STOP);
2093 	CSR_WRITE_4(sc, VGE_ISR, 0xFFFFFFFF);
2094 	CSR_WRITE_2(sc, VGE_TXQCSRC, 0xFFFF);
2095 	CSR_WRITE_1(sc, VGE_RXQCSRC, 0xFF);
2096 	CSR_WRITE_4(sc, VGE_RXDESC_ADDR_LO, 0);
2097 
2098 	if (sc->sc_rx_mhead != NULL) {
2099 		m_freem(sc->sc_rx_mhead);
2100 		sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
2101 	}
2102 
2103 	/* Free the TX list buffers. */
2104 
2105 	for (i = 0; i < VGE_NTXDESC; i++) {
2106 		txs = &sc->sc_txsoft[i];
2107 		if (txs->txs_mbuf != NULL) {
2108 			bus_dmamap_unload(sc->sc_dmat, txs->txs_dmamap);
2109 			m_freem(txs->txs_mbuf);
2110 			txs->txs_mbuf = NULL;
2111 		}
2112 	}
2113 
2114 	/* Free the RX list buffers. */
2115 
2116 	for (i = 0; i < VGE_NRXDESC; i++) {
2117 		rxs = &sc->sc_rxsoft[i];
2118 		if (rxs->rxs_mbuf != NULL) {
2119 			bus_dmamap_unload(sc->sc_dmat, rxs->rxs_dmamap);
2120 			m_freem(rxs->rxs_mbuf);
2121 			rxs->rxs_mbuf = NULL;
2122 		}
2123 	}
2124 
2125 	splx(s);
2126 }
2127 
2128 #if VGE_POWER_MANAGEMENT
2129 /*
2130  * Device suspend routine.  Stop the interface and save some PCI
2131  * settings in case the BIOS doesn't restore them properly on
2132  * resume.
2133  */
2134 static int
2135 vge_suspend(struct device *dev)
2136 {
2137 	struct vge_softc *sc;
2138 	int i;
2139 
2140 	sc = device_get_softc(dev);
2141 
2142 	vge_stop(sc);
2143 
2144         for (i = 0; i < 5; i++)
2145 		sc->sc_saved_maps[i] =
2146 		    pci_read_config(dev, PCIR_MAPS + i * 4, 4);
2147 	sc->sc_saved_biosaddr = pci_read_config(dev, PCIR_BIOS, 4);
2148 	sc->sc_saved_intline = pci_read_config(dev, PCIR_INTLINE, 1);
2149 	sc->sc_saved_cachelnsz = pci_read_config(dev, PCIR_CACHELNSZ, 1);
2150 	sc->sc_saved_lattimer = pci_read_config(dev, PCIR_LATTIMER, 1);
2151 
2152 	sc->suspended = 1;
2153 
2154 	return 0;
2155 }
2156 
2157 /*
2158  * Device resume routine.  Restore some PCI settings in case the BIOS
2159  * doesn't, re-enable busmastering, and restart the interface if
2160  * appropriate.
2161  */
2162 static int
2163 vge_resume(struct device *dev)
2164 {
2165 	struct vge_softc *sc;
2166 	struct ifnet *ifp;
2167 	int i;
2168 
2169 	sc = (void *)dev;
2170 	ifp = &sc->sc_ethercom.ec_if;
2171 
2172         /* better way to do this? */
2173 	for (i = 0; i < 5; i++)
2174 		pci_write_config(dev, PCIR_MAPS + i * 4,
2175 		    sc->sc_saved_maps[i], 4);
2176 	pci_write_config(dev, PCIR_BIOS, sc->sc_saved_biosaddr, 4);
2177 	pci_write_config(dev, PCIR_INTLINE, sc->sc_saved_intline, 1);
2178 	pci_write_config(dev, PCIR_CACHELNSZ, sc->sc_saved_cachelnsz, 1);
2179 	pci_write_config(dev, PCIR_LATTIMER, sc->sc_saved_lattimer, 1);
2180 
2181 	/* reenable busmastering */
2182 	pci_enable_busmaster(dev);
2183 	pci_enable_io(dev, SYS_RES_MEMORY);
2184 
2185 	/* reinitialize interface if necessary */
2186 	if (ifp->if_flags & IFF_UP)
2187 		vge_init(sc);
2188 
2189 	sc->suspended = 0;
2190 
2191 	return 0;
2192 }
2193 #endif
2194 
2195 /*
2196  * Stop all chip I/O so that the kernel's probe routines don't
2197  * get confused by errant DMAs when rebooting.
2198  */
2199 static void
2200 vge_shutdown(void *arg)
2201 {
2202 	struct vge_softc *sc;
2203 
2204 	sc = arg;
2205 	vge_stop(&sc->sc_ethercom.ec_if, 1);
2206 }
2207