xref: /netbsd-src/sys/dev/pci/if_vge.c (revision cac8e449158efc7261bebc8657cbb0125a2cfdde)
1 /* $NetBSD: if_vge.c,v 1.41 2008/04/10 19:13:37 cegger 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 1.41 2008/04/10 19:13:37 cegger 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_JUMBO_MTU		9000
119 
120 #define VGE_IFQ_MAXLEN		64
121 
122 #define VGE_RING_ALIGN		256
123 
124 #define VGE_NTXDESC		256
125 #define VGE_NTXDESC_MASK	(VGE_NTXDESC - 1)
126 #define VGE_NEXT_TXDESC(x)	((x + 1) & VGE_NTXDESC_MASK)
127 #define VGE_PREV_TXDESC(x)	((x - 1) & VGE_NTXDESC_MASK)
128 
129 #define VGE_NRXDESC		256	/* Must be a multiple of 4!! */
130 #define VGE_NRXDESC_MASK	(VGE_NRXDESC - 1)
131 #define VGE_NEXT_RXDESC(x)	((x + 1) & VGE_NRXDESC_MASK)
132 #define VGE_PREV_RXDESC(x)	((x - 1) & VGE_NRXDESC_MASK)
133 
134 #define VGE_ADDR_LO(y)		((uint64_t)(y) & 0xFFFFFFFF)
135 #define VGE_ADDR_HI(y)		((uint64_t)(y) >> 32)
136 #define VGE_BUFLEN(y)		((y) & 0x7FFF)
137 #define ETHER_PAD_LEN		(ETHER_MIN_LEN - ETHER_CRC_LEN)
138 
139 #define VGE_POWER_MANAGEMENT	0	/* disabled for now */
140 
141 /*
142  * Mbuf adjust factor to force 32-bit alignment of IP header.
143  * Drivers should pad ETHER_ALIGN bytes when setting up a
144  * RX mbuf so the upper layers get the IP header properly aligned
145  * past the 14-byte Ethernet header.
146  *
147  * See also comment in vge_encap().
148  */
149 #define ETHER_ALIGN		2
150 
151 #ifdef __NO_STRICT_ALIGNMENT
152 #define VGE_RX_BUFSIZE		MCLBYTES
153 #else
154 #define VGE_RX_PAD		sizeof(uint32_t)
155 #define VGE_RX_BUFSIZE		(MCLBYTES - VGE_RX_PAD)
156 #endif
157 
158 /*
159  * Control structures are DMA'd to the vge chip. We allocate them in
160  * a single clump that maps to a single DMA segment to make several things
161  * easier.
162  */
163 struct vge_control_data {
164 	/* TX descriptors */
165 	struct vge_txdesc	vcd_txdescs[VGE_NTXDESC];
166 	/* RX descriptors */
167 	struct vge_rxdesc	vcd_rxdescs[VGE_NRXDESC];
168 	/* dummy data for TX padding */
169 	uint8_t			vcd_pad[ETHER_PAD_LEN];
170 };
171 
172 #define VGE_CDOFF(x)	offsetof(struct vge_control_data, x)
173 #define VGE_CDTXOFF(x)	VGE_CDOFF(vcd_txdescs[(x)])
174 #define VGE_CDRXOFF(x)	VGE_CDOFF(vcd_rxdescs[(x)])
175 #define VGE_CDPADOFF()	VGE_CDOFF(vcd_pad[0])
176 
177 /*
178  * Software state for TX jobs.
179  */
180 struct vge_txsoft {
181 	struct mbuf	*txs_mbuf;		/* head of our mbuf chain */
182 	bus_dmamap_t	txs_dmamap;		/* our DMA map */
183 };
184 
185 /*
186  * Software state for RX jobs.
187  */
188 struct vge_rxsoft {
189 	struct mbuf	*rxs_mbuf;		/* head of our mbuf chain */
190 	bus_dmamap_t	rxs_dmamap;		/* our DMA map */
191 };
192 
193 
194 struct vge_softc {
195 	struct device		sc_dev;
196 
197 	bus_space_tag_t		sc_bst;		/* bus space tag */
198 	bus_space_handle_t	sc_bsh;		/* bus space handle */
199 	bus_dma_tag_t		sc_dmat;
200 
201 	struct ethercom		sc_ethercom;	/* interface info */
202 	uint8_t			sc_eaddr[ETHER_ADDR_LEN];
203 
204 	void			*sc_intrhand;
205 	struct mii_data		sc_mii;
206 	uint8_t			sc_type;
207 	int			sc_if_flags;
208 	int			sc_link;
209 	int			sc_camidx;
210 	callout_t		sc_timeout;
211 
212 	bus_dmamap_t		sc_cddmamap;
213 #define sc_cddma		sc_cddmamap->dm_segs[0].ds_addr
214 
215 	struct vge_txsoft	sc_txsoft[VGE_NTXDESC];
216 	struct vge_rxsoft	sc_rxsoft[VGE_NRXDESC];
217 	struct vge_control_data	*sc_control_data;
218 #define sc_txdescs		sc_control_data->vcd_txdescs
219 #define sc_rxdescs		sc_control_data->vcd_rxdescs
220 
221 	int			sc_tx_prodidx;
222 	int			sc_tx_considx;
223 	int			sc_tx_free;
224 
225 	struct mbuf		*sc_rx_mhead;
226 	struct mbuf		*sc_rx_mtail;
227 	int			sc_rx_prodidx;
228 	int			sc_rx_consumed;
229 
230 	int			sc_suspended;	/* 0 = normal  1 = suspended */
231 	uint32_t		sc_saved_maps[5];	/* pci data */
232 	uint32_t		sc_saved_biosaddr;
233 	uint8_t			sc_saved_intline;
234 	uint8_t			sc_saved_cachelnsz;
235 	uint8_t			sc_saved_lattimer;
236 };
237 
238 #define VGE_CDTXADDR(sc, x)	((sc)->sc_cddma + VGE_CDTXOFF(x))
239 #define VGE_CDRXADDR(sc, x)	((sc)->sc_cddma + VGE_CDRXOFF(x))
240 #define VGE_CDPADADDR(sc)	((sc)->sc_cddma + VGE_CDPADOFF())
241 
242 #define VGE_TXDESCSYNC(sc, idx, ops)					\
243 	bus_dmamap_sync((sc)->sc_dmat,(sc)->sc_cddmamap,		\
244 	    VGE_CDTXOFF(idx),						\
245 	    offsetof(struct vge_txdesc, td_frag[0]),			\
246 	    (ops))
247 #define VGE_TXFRAGSYNC(sc, idx, nsegs, ops)				\
248 	bus_dmamap_sync((sc)->sc_dmat, (sc)->sc_cddmamap,		\
249 	    VGE_CDTXOFF(idx) +						\
250 	    offsetof(struct vge_txdesc, td_frag[0]),			\
251 	    sizeof(struct vge_txfrag) * (nsegs),			\
252 	    (ops))
253 #define VGE_RXDESCSYNC(sc, idx, ops)					\
254 	bus_dmamap_sync((sc)->sc_dmat, (sc)->sc_cddmamap,		\
255 	    VGE_CDRXOFF(idx),						\
256 	    sizeof(struct vge_rxdesc),					\
257 	    (ops))
258 
259 /*
260  * register space access macros
261  */
262 #define CSR_WRITE_4(sc, reg, val)	\
263 	bus_space_write_4((sc)->sc_bst, (sc)->sc_bsh, (reg), (val))
264 #define CSR_WRITE_2(sc, reg, val)	\
265 	bus_space_write_2((sc)->sc_bst, (sc)->sc_bsh, (reg), (val))
266 #define CSR_WRITE_1(sc, reg, val)	\
267 	bus_space_write_1((sc)->sc_bst, (sc)->sc_bsh, (reg), (val))
268 
269 #define CSR_READ_4(sc, reg)		\
270 	bus_space_read_4((sc)->sc_bst, (sc)->sc_bsh, (reg))
271 #define CSR_READ_2(sc, reg)		\
272 	bus_space_read_2((sc)->sc_bst, (sc)->sc_bsh, (reg))
273 #define CSR_READ_1(sc, reg)		\
274 	bus_space_read_1((sc)->sc_bst, (sc)->sc_bsh, (reg))
275 
276 #define CSR_SETBIT_1(sc, reg, x)	\
277 	CSR_WRITE_1((sc), (reg), CSR_READ_1((sc), (reg)) | (x))
278 #define CSR_SETBIT_2(sc, reg, x)	\
279 	CSR_WRITE_2((sc), (reg), CSR_READ_2((sc), (reg)) | (x))
280 #define CSR_SETBIT_4(sc, reg, x)	\
281 	CSR_WRITE_4((sc), (reg), CSR_READ_4((sc), (reg)) | (x))
282 
283 #define CSR_CLRBIT_1(sc, reg, x)	\
284 	CSR_WRITE_1((sc), (reg), CSR_READ_1((sc), (reg)) & ~(x))
285 #define CSR_CLRBIT_2(sc, reg, x)	\
286 	CSR_WRITE_2((sc), (reg), CSR_READ_2((sc), (reg)) & ~(x))
287 #define CSR_CLRBIT_4(sc, reg, x)	\
288 	CSR_WRITE_4((sc), (reg), CSR_READ_4((sc), (reg)) & ~(x))
289 
290 #define VGE_TIMEOUT		10000
291 
292 #define VGE_PCI_LOIO             0x10
293 #define VGE_PCI_LOMEM            0x14
294 
295 static inline void vge_set_txaddr(struct vge_txfrag *, bus_addr_t);
296 static inline void vge_set_rxaddr(struct vge_rxdesc *, bus_addr_t);
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 vge_softc *);
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 
1008 	/*
1009 	 * We can support 802.1Q VLAN-sized frames and jumbo
1010 	 * Ethernet frames.
1011 	 */
1012 	sc->sc_ethercom.ec_capabilities |=
1013 	    ETHERCAP_VLAN_MTU | ETHERCAP_JUMBO_MTU |
1014 	    ETHERCAP_VLAN_HWTAGGING;
1015 
1016 	/*
1017 	 * We can do IPv4/TCPv4/UDPv4 checksums in hardware.
1018 	 */
1019 	ifp->if_capabilities |=
1020 	    IFCAP_CSUM_IPv4_Tx | IFCAP_CSUM_IPv4_Rx |
1021 	    IFCAP_CSUM_TCPv4_Tx | IFCAP_CSUM_TCPv4_Rx |
1022 	    IFCAP_CSUM_UDPv4_Tx | IFCAP_CSUM_UDPv4_Rx;
1023 
1024 #ifdef DEVICE_POLLING
1025 #ifdef IFCAP_POLLING
1026 	ifp->if_capabilities |= IFCAP_POLLING;
1027 #endif
1028 #endif
1029 	ifp->if_watchdog = vge_watchdog;
1030 	ifp->if_init = vge_init;
1031 	IFQ_SET_MAXLEN(&ifp->if_snd, max(VGE_IFQ_MAXLEN, IFQ_MAXLEN));
1032 
1033 	/*
1034 	 * Initialize our media structures and probe the MII.
1035 	 */
1036 	sc->sc_mii.mii_ifp = ifp;
1037 	sc->sc_mii.mii_readreg = vge_miibus_readreg;
1038 	sc->sc_mii.mii_writereg = vge_miibus_writereg;
1039 	sc->sc_mii.mii_statchg = vge_miibus_statchg;
1040 
1041 	sc->sc_ethercom.ec_mii = &sc->sc_mii;
1042 	ifmedia_init(&sc->sc_mii.mii_media, 0, ether_mediachange,
1043 	    ether_mediastatus);
1044 	mii_attach(&sc->sc_dev, &sc->sc_mii, 0xffffffff, MII_PHY_ANY,
1045 	    MII_OFFSET_ANY, MIIF_DOPAUSE);
1046 	if (LIST_FIRST(&sc->sc_mii.mii_phys) == NULL) {
1047 		ifmedia_add(&sc->sc_mii.mii_media, IFM_ETHER|IFM_NONE, 0, NULL);
1048 		ifmedia_set(&sc->sc_mii.mii_media, IFM_ETHER|IFM_NONE);
1049 	} else
1050 		ifmedia_set(&sc->sc_mii.mii_media, IFM_ETHER|IFM_AUTO);
1051 
1052 	/*
1053 	 * Attach the interface.
1054 	 */
1055 	if_attach(ifp);
1056 	ether_ifattach(ifp, eaddr);
1057 
1058 	callout_init(&sc->sc_timeout, 0);
1059 	callout_setfunc(&sc->sc_timeout, vge_tick, sc);
1060 
1061 	/*
1062 	 * Make sure the interface is shutdown during reboot.
1063 	 */
1064 	if (shutdownhook_establish(vge_shutdown, sc) == NULL) {
1065 		aprint_error_dev(&sc->sc_dev, "WARNING: unable to establish shutdown hook\n");
1066 	}
1067 }
1068 
1069 static int
1070 vge_newbuf(struct vge_softc *sc, int idx, struct mbuf *m)
1071 {
1072 	struct mbuf *m_new;
1073 	struct vge_rxdesc *rxd;
1074 	struct vge_rxsoft *rxs;
1075 	bus_dmamap_t map;
1076 	int i;
1077 #ifdef DIAGNOSTIC
1078 	uint32_t rd_sts;
1079 #endif
1080 
1081 	m_new = NULL;
1082 	if (m == NULL) {
1083 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
1084 		if (m_new == NULL)
1085 			return ENOBUFS;
1086 
1087 		MCLGET(m_new, M_DONTWAIT);
1088 		if ((m_new->m_flags & M_EXT) == 0) {
1089 			m_freem(m_new);
1090 			return ENOBUFS;
1091 		}
1092 
1093 		m = m_new;
1094 	} else
1095 		m->m_data = m->m_ext.ext_buf;
1096 
1097 
1098 	/*
1099 	 * This is part of an evil trick to deal with non-x86 platforms.
1100 	 * The VIA chip requires RX buffers to be aligned on 32-bit
1101 	 * boundaries, but that will hose non-x86 machines. To get around
1102 	 * this, we leave some empty space at the start of each buffer
1103 	 * and for non-x86 hosts, we copy the buffer back two bytes
1104 	 * to achieve word alignment. This is slightly more efficient
1105 	 * than allocating a new buffer, copying the contents, and
1106 	 * discarding the old buffer.
1107 	 */
1108 	m->m_len = m->m_pkthdr.len = VGE_RX_BUFSIZE;
1109 #ifndef __NO_STRICT_ALIGNMENT
1110 	m->m_data += VGE_RX_PAD;
1111 #endif
1112 	rxs = &sc->sc_rxsoft[idx];
1113 	map = rxs->rxs_dmamap;
1114 
1115 	if (bus_dmamap_load_mbuf(sc->sc_dmat, map, m, BUS_DMA_NOWAIT) != 0)
1116 		goto out;
1117 
1118 	rxd = &sc->sc_rxdescs[idx];
1119 
1120 #ifdef DIAGNOSTIC
1121 	/* If this descriptor is still owned by the chip, bail. */
1122 	VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1123 	rd_sts = le32toh(rxd->rd_sts);
1124 	VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1125 	if (rd_sts & VGE_RDSTS_OWN) {
1126 		panic("%s: tried to map busy RX descriptor",
1127 		    device_xname(&sc->sc_dev));
1128 	}
1129 #endif
1130 
1131 	rxs->rxs_mbuf = m;
1132 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1133 	    BUS_DMASYNC_PREREAD);
1134 
1135 	rxd->rd_buflen =
1136 	    htole16(VGE_BUFLEN(map->dm_segs[0].ds_len) | VGE_RXDESC_I);
1137 	vge_set_rxaddr(rxd, map->dm_segs[0].ds_addr);
1138 	rxd->rd_sts = 0;
1139 	rxd->rd_ctl = 0;
1140 	VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1141 
1142 	/*
1143 	 * Note: the manual fails to document the fact that for
1144 	 * proper opration, the driver needs to replentish the RX
1145 	 * DMA ring 4 descriptors at a time (rather than one at a
1146 	 * time, like most chips). We can allocate the new buffers
1147 	 * but we should not set the OWN bits until we're ready
1148 	 * to hand back 4 of them in one shot.
1149 	 */
1150 
1151 #define VGE_RXCHUNK 4
1152 	sc->sc_rx_consumed++;
1153 	if (sc->sc_rx_consumed == VGE_RXCHUNK) {
1154 		for (i = idx; i != idx - VGE_RXCHUNK; i--) {
1155 			KASSERT(i >= 0);
1156 			sc->sc_rxdescs[i].rd_sts |= htole32(VGE_RDSTS_OWN);
1157 			VGE_RXDESCSYNC(sc, i,
1158 			    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1159 		}
1160 		sc->sc_rx_consumed = 0;
1161 	}
1162 
1163 	return 0;
1164  out:
1165 	if (m_new != NULL)
1166 		m_freem(m_new);
1167 	return ENOMEM;
1168 }
1169 
1170 #ifndef __NO_STRICT_ALIGNMENT
1171 static inline void
1172 vge_fixup_rx(struct mbuf *m)
1173 {
1174 	int i;
1175 	uint16_t *src, *dst;
1176 
1177 	src = mtod(m, uint16_t *);
1178 	dst = src - 1;
1179 
1180 	for (i = 0; i < (m->m_len / sizeof(uint16_t) + 1); i++)
1181 		*dst++ = *src++;
1182 
1183 	m->m_data -= ETHER_ALIGN;
1184 }
1185 #endif
1186 
1187 /*
1188  * RX handler. We support the reception of jumbo frames that have
1189  * been fragmented across multiple 2K mbuf cluster buffers.
1190  */
1191 static void
1192 vge_rxeof(struct vge_softc *sc)
1193 {
1194 	struct mbuf *m;
1195 	struct ifnet *ifp;
1196 	int idx, total_len, lim;
1197 	struct vge_rxdesc *cur_rxd;
1198 	struct vge_rxsoft *rxs;
1199 	uint32_t rxstat, rxctl;
1200 
1201 	ifp = &sc->sc_ethercom.ec_if;
1202 	lim = 0;
1203 
1204 	/* Invalidate the descriptor memory */
1205 
1206 	for (idx = sc->sc_rx_prodidx;; idx = VGE_NEXT_RXDESC(idx)) {
1207 		cur_rxd = &sc->sc_rxdescs[idx];
1208 
1209 		VGE_RXDESCSYNC(sc, idx,
1210 		    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1211 		rxstat = le32toh(cur_rxd->rd_sts);
1212 		if ((rxstat & VGE_RDSTS_OWN) != 0) {
1213 			VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1214 			break;
1215 		}
1216 
1217 		rxctl = le32toh(cur_rxd->rd_ctl);
1218 		rxs = &sc->sc_rxsoft[idx];
1219 		m = rxs->rxs_mbuf;
1220 		total_len = (rxstat & VGE_RDSTS_BUFSIZ) >> 16;
1221 
1222 		/* Invalidate the RX mbuf and unload its map */
1223 
1224 		bus_dmamap_sync(sc->sc_dmat, rxs->rxs_dmamap,
1225 		    0, rxs->rxs_dmamap->dm_mapsize, BUS_DMASYNC_POSTREAD);
1226 		bus_dmamap_unload(sc->sc_dmat, rxs->rxs_dmamap);
1227 
1228 		/*
1229 		 * If the 'start of frame' bit is set, this indicates
1230 		 * either the first fragment in a multi-fragment receive,
1231 		 * or an intermediate fragment. Either way, we want to
1232 		 * accumulate the buffers.
1233 		 */
1234 		if (rxstat & VGE_RXPKT_SOF) {
1235 			m->m_len = VGE_RX_BUFSIZE;
1236 			if (sc->sc_rx_mhead == NULL)
1237 				sc->sc_rx_mhead = sc->sc_rx_mtail = m;
1238 			else {
1239 				m->m_flags &= ~M_PKTHDR;
1240 				sc->sc_rx_mtail->m_next = m;
1241 				sc->sc_rx_mtail = m;
1242 			}
1243 			vge_newbuf(sc, idx, NULL);
1244 			continue;
1245 		}
1246 
1247 		/*
1248 		 * Bad/error frames will have the RXOK bit cleared.
1249 		 * However, there's one error case we want to allow:
1250 		 * if a VLAN tagged frame arrives and the chip can't
1251 		 * match it against the CAM filter, it considers this
1252 		 * a 'VLAN CAM filter miss' and clears the 'RXOK' bit.
1253 		 * We don't want to drop the frame though: our VLAN
1254 		 * filtering is done in software.
1255 		 */
1256 		if ((rxstat & VGE_RDSTS_RXOK) == 0 &&
1257 		    (rxstat & VGE_RDSTS_VIDM) == 0 &&
1258 		    (rxstat & VGE_RDSTS_CSUMERR) == 0) {
1259 			ifp->if_ierrors++;
1260 			/*
1261 			 * If this is part of a multi-fragment packet,
1262 			 * discard all the pieces.
1263 			 */
1264 			if (sc->sc_rx_mhead != NULL) {
1265 				m_freem(sc->sc_rx_mhead);
1266 				sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1267 			}
1268 			vge_newbuf(sc, idx, m);
1269 			continue;
1270 		}
1271 
1272 		/*
1273 		 * If allocating a replacement mbuf fails,
1274 		 * reload the current one.
1275 		 */
1276 
1277 		if (vge_newbuf(sc, idx, NULL)) {
1278 			ifp->if_ierrors++;
1279 			if (sc->sc_rx_mhead != NULL) {
1280 				m_freem(sc->sc_rx_mhead);
1281 				sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1282 			}
1283 			vge_newbuf(sc, idx, m);
1284 			continue;
1285 		}
1286 
1287 		if (sc->sc_rx_mhead != NULL) {
1288 			m->m_len = total_len % VGE_RX_BUFSIZE;
1289 			/*
1290 			 * Special case: if there's 4 bytes or less
1291 			 * in this buffer, the mbuf can be discarded:
1292 			 * the last 4 bytes is the CRC, which we don't
1293 			 * care about anyway.
1294 			 */
1295 			if (m->m_len <= ETHER_CRC_LEN) {
1296 				sc->sc_rx_mtail->m_len -=
1297 				    (ETHER_CRC_LEN - m->m_len);
1298 				m_freem(m);
1299 			} else {
1300 				m->m_len -= ETHER_CRC_LEN;
1301 				m->m_flags &= ~M_PKTHDR;
1302 				sc->sc_rx_mtail->m_next = m;
1303 			}
1304 			m = sc->sc_rx_mhead;
1305 			sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1306 			m->m_pkthdr.len = total_len - ETHER_CRC_LEN;
1307 		} else
1308 			m->m_pkthdr.len = m->m_len = total_len - ETHER_CRC_LEN;
1309 
1310 #ifndef __NO_STRICT_ALIGNMENT
1311 		vge_fixup_rx(m);
1312 #endif
1313 		ifp->if_ipackets++;
1314 		m->m_pkthdr.rcvif = ifp;
1315 
1316 		/* Do RX checksumming if enabled */
1317 		if (ifp->if_csum_flags_rx & M_CSUM_IPv4) {
1318 
1319 			/* Check IP header checksum */
1320 			if (rxctl & VGE_RDCTL_IPPKT)
1321 				m->m_pkthdr.csum_flags |= M_CSUM_IPv4;
1322 			if ((rxctl & VGE_RDCTL_IPCSUMOK) == 0)
1323 				m->m_pkthdr.csum_flags |= M_CSUM_IPv4_BAD;
1324 		}
1325 
1326 		if (ifp->if_csum_flags_rx & M_CSUM_TCPv4) {
1327 			/* Check UDP checksum */
1328 			if (rxctl & VGE_RDCTL_TCPPKT)
1329 				m->m_pkthdr.csum_flags |= M_CSUM_TCPv4;
1330 
1331 			if ((rxctl & VGE_RDCTL_PROTOCSUMOK) == 0)
1332 				m->m_pkthdr.csum_flags |= M_CSUM_TCP_UDP_BAD;
1333 		}
1334 
1335 		if (ifp->if_csum_flags_rx & M_CSUM_UDPv4) {
1336 			/* Check UDP checksum */
1337 			if (rxctl & VGE_RDCTL_UDPPKT)
1338 				m->m_pkthdr.csum_flags |= M_CSUM_UDPv4;
1339 
1340 			if ((rxctl & VGE_RDCTL_PROTOCSUMOK) == 0)
1341 				m->m_pkthdr.csum_flags |= M_CSUM_TCP_UDP_BAD;
1342 		}
1343 
1344 		if (rxstat & VGE_RDSTS_VTAG) {
1345 			/*
1346 			 * We use bswap16() here because:
1347 			 * On LE machines, tag is stored in BE as stream data.
1348 			 * On BE machines, tag is stored in BE as stream data
1349 			 *  but it was already swapped by le32toh() above.
1350 			 */
1351 			VLAN_INPUT_TAG(ifp, m,
1352 			    bswap16(rxctl & VGE_RDCTL_VLANID), continue);
1353 		}
1354 
1355 #if NBPFILTER > 0
1356 		/*
1357 		 * Handle BPF listeners.
1358 		 */
1359 		if (ifp->if_bpf)
1360 			bpf_mtap(ifp->if_bpf, m);
1361 #endif
1362 
1363 		(*ifp->if_input)(ifp, m);
1364 
1365 		lim++;
1366 		if (lim == VGE_NRXDESC)
1367 			break;
1368 	}
1369 
1370 	sc->sc_rx_prodidx = idx;
1371 	CSR_WRITE_2(sc, VGE_RXDESC_RESIDUECNT, lim);
1372 }
1373 
1374 static void
1375 vge_txeof(struct vge_softc *sc)
1376 {
1377 	struct ifnet *ifp;
1378 	struct vge_txsoft *txs;
1379 	uint32_t txstat;
1380 	int idx;
1381 
1382 	ifp = &sc->sc_ethercom.ec_if;
1383 
1384 	for (idx = sc->sc_tx_considx;
1385 	    sc->sc_tx_free < VGE_NTXDESC;
1386 	    idx = VGE_NEXT_TXDESC(idx), sc->sc_tx_free++) {
1387 		VGE_TXDESCSYNC(sc, idx,
1388 		    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1389 		txstat = le32toh(sc->sc_txdescs[idx].td_sts);
1390 		VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1391 		if (txstat & VGE_TDSTS_OWN) {
1392 			break;
1393 		}
1394 
1395 		txs = &sc->sc_txsoft[idx];
1396 		m_freem(txs->txs_mbuf);
1397 		txs->txs_mbuf = NULL;
1398 		bus_dmamap_sync(sc->sc_dmat, txs->txs_dmamap, 0,
1399 		    txs->txs_dmamap->dm_mapsize, BUS_DMASYNC_POSTWRITE);
1400 		bus_dmamap_unload(sc->sc_dmat, txs->txs_dmamap);
1401 		if (txstat & (VGE_TDSTS_EXCESSCOLL|VGE_TDSTS_COLL))
1402 			ifp->if_collisions++;
1403 		if (txstat & VGE_TDSTS_TXERR)
1404 			ifp->if_oerrors++;
1405 		else
1406 			ifp->if_opackets++;
1407 	}
1408 
1409 	sc->sc_tx_considx = idx;
1410 
1411 	if (sc->sc_tx_free > 0) {
1412 		ifp->if_flags &= ~IFF_OACTIVE;
1413 	}
1414 
1415 	/*
1416 	 * If not all descriptors have been released reaped yet,
1417 	 * reload the timer so that we will eventually get another
1418 	 * interrupt that will cause us to re-enter this routine.
1419 	 * This is done in case the transmitter has gone idle.
1420 	 */
1421 	if (sc->sc_tx_free < VGE_NTXDESC)
1422 		CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_TIMER0_ENABLE);
1423 	else
1424 		ifp->if_timer = 0;
1425 }
1426 
1427 static void
1428 vge_tick(void *xsc)
1429 {
1430 	struct vge_softc *sc;
1431 	struct ifnet *ifp;
1432 	struct mii_data *mii;
1433 	int s;
1434 
1435 	sc = xsc;
1436 	ifp = &sc->sc_ethercom.ec_if;
1437 	mii = &sc->sc_mii;
1438 
1439 	s = splnet();
1440 
1441 	callout_schedule(&sc->sc_timeout, hz);
1442 
1443 	mii_tick(mii);
1444 	if (sc->sc_link) {
1445 		if ((mii->mii_media_status & IFM_ACTIVE) == 0)
1446 			sc->sc_link = 0;
1447 	} else {
1448 		if (mii->mii_media_status & IFM_ACTIVE &&
1449 		    IFM_SUBTYPE(mii->mii_media_active) != IFM_NONE) {
1450 			sc->sc_link = 1;
1451 			if (!IFQ_IS_EMPTY(&ifp->if_snd))
1452 				vge_start(ifp);
1453 		}
1454 	}
1455 
1456 	splx(s);
1457 }
1458 
1459 static int
1460 vge_intr(void *arg)
1461 {
1462 	struct vge_softc *sc;
1463 	struct ifnet *ifp;
1464 	uint32_t status;
1465 	int claim;
1466 
1467 	sc = arg;
1468 	claim = 0;
1469 	if (sc->sc_suspended) {
1470 		return claim;
1471 	}
1472 
1473 	ifp = &sc->sc_ethercom.ec_if;
1474 
1475 	if ((ifp->if_flags & IFF_UP) == 0) {
1476 		return claim;
1477 	}
1478 
1479 	/* Disable interrupts */
1480 	CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
1481 
1482 	for (;;) {
1483 
1484 		status = CSR_READ_4(sc, VGE_ISR);
1485 		/* If the card has gone away the read returns 0xffff. */
1486 		if (status == 0xFFFFFFFF)
1487 			break;
1488 
1489 		if (status) {
1490 			claim = 1;
1491 			CSR_WRITE_4(sc, VGE_ISR, status);
1492 		}
1493 
1494 		if ((status & VGE_INTRS) == 0)
1495 			break;
1496 
1497 		if (status & (VGE_ISR_RXOK|VGE_ISR_RXOK_HIPRIO))
1498 			vge_rxeof(sc);
1499 
1500 		if (status & (VGE_ISR_RXOFLOW|VGE_ISR_RXNODESC)) {
1501 			vge_rxeof(sc);
1502 			CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_RUN);
1503 			CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_WAK);
1504 		}
1505 
1506 		if (status & (VGE_ISR_TXOK0|VGE_ISR_TIMER0))
1507 			vge_txeof(sc);
1508 
1509 		if (status & (VGE_ISR_TXDMA_STALL|VGE_ISR_RXDMA_STALL))
1510 			vge_init(ifp);
1511 
1512 		if (status & VGE_ISR_LINKSTS)
1513 			vge_tick(sc);
1514 	}
1515 
1516 	/* Re-enable interrupts */
1517 	CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_GMSK);
1518 
1519 	if (claim && !IFQ_IS_EMPTY(&ifp->if_snd))
1520 		vge_start(ifp);
1521 
1522 	return claim;
1523 }
1524 
1525 static int
1526 vge_encap(struct vge_softc *sc, struct mbuf *m_head, int idx)
1527 {
1528 	struct vge_txsoft *txs;
1529 	struct vge_txdesc *txd;
1530 	struct vge_txfrag *f;
1531 	struct mbuf *m_new;
1532 	bus_dmamap_t map;
1533 	int m_csumflags, seg, error, flags;
1534 	struct m_tag *mtag;
1535 	size_t sz;
1536 	uint32_t td_sts, td_ctl;
1537 
1538 	KASSERT(sc->sc_tx_free > 0);
1539 
1540 	txd = &sc->sc_txdescs[idx];
1541 
1542 #ifdef DIAGNOSTIC
1543 	/* If this descriptor is still owned by the chip, bail. */
1544 	VGE_TXDESCSYNC(sc, idx,
1545 	    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1546 	td_sts = le32toh(txd->td_sts);
1547 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1548 	if (td_sts & VGE_TDSTS_OWN) {
1549 		return ENOBUFS;
1550 	}
1551 #endif
1552 
1553 	/*
1554 	 * Preserve m_pkthdr.csum_flags here since m_head might be
1555 	 * updated by m_defrag()
1556 	 */
1557 	m_csumflags = m_head->m_pkthdr.csum_flags;
1558 
1559 	txs = &sc->sc_txsoft[idx];
1560 	map = txs->txs_dmamap;
1561 	error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m_head, BUS_DMA_NOWAIT);
1562 
1563 	/* If too many segments to map, coalesce */
1564 	if (error == EFBIG ||
1565 	    (m_head->m_pkthdr.len < ETHER_PAD_LEN &&
1566 	     map->dm_nsegs == VGE_TX_FRAGS)) {
1567 		m_new = m_defrag(m_head, M_DONTWAIT);
1568 		if (m_new == NULL)
1569 			return EFBIG;
1570 
1571 		error = bus_dmamap_load_mbuf(sc->sc_dmat, map,
1572 		    m_new, BUS_DMA_NOWAIT);
1573 		if (error) {
1574 			m_freem(m_new);
1575 			return error;
1576 		}
1577 
1578 		m_head = m_new;
1579 	} else if (error)
1580 		return error;
1581 
1582 	txs->txs_mbuf = m_head;
1583 
1584 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1585 	    BUS_DMASYNC_PREWRITE);
1586 
1587 	for (seg = 0, f = &txd->td_frag[0]; seg < map->dm_nsegs; seg++, f++) {
1588 		f->tf_buflen = htole16(VGE_BUFLEN(map->dm_segs[seg].ds_len));
1589 		vge_set_txaddr(f, map->dm_segs[seg].ds_addr);
1590 	}
1591 
1592 	/* Argh. This chip does not autopad short frames */
1593 	sz = m_head->m_pkthdr.len;
1594 	if (sz < ETHER_PAD_LEN) {
1595 		f->tf_buflen = htole16(VGE_BUFLEN(ETHER_PAD_LEN - sz));
1596 		vge_set_txaddr(f, VGE_CDPADADDR(sc));
1597 		sz = ETHER_PAD_LEN;
1598 		seg++;
1599 	}
1600 	VGE_TXFRAGSYNC(sc, idx, seg, BUS_DMASYNC_PREWRITE);
1601 
1602 	/*
1603 	 * When telling the chip how many segments there are, we
1604 	 * must use nsegs + 1 instead of just nsegs. Darned if I
1605 	 * know why.
1606 	 */
1607 	seg++;
1608 
1609 	flags = 0;
1610 	if (m_csumflags & M_CSUM_IPv4)
1611 		flags |= VGE_TDCTL_IPCSUM;
1612 	if (m_csumflags & M_CSUM_TCPv4)
1613 		flags |= VGE_TDCTL_TCPCSUM;
1614 	if (m_csumflags & M_CSUM_UDPv4)
1615 		flags |= VGE_TDCTL_UDPCSUM;
1616 	td_sts = sz << 16;
1617 	td_ctl = flags | (seg << 28) | VGE_TD_LS_NORM;
1618 
1619 	if (sz > ETHERMTU + ETHER_HDR_LEN)
1620 		td_ctl |= VGE_TDCTL_JUMBO;
1621 
1622 	/*
1623 	 * Set up hardware VLAN tagging.
1624 	 */
1625 	mtag = VLAN_OUTPUT_TAG(&sc->sc_ethercom, m_head);
1626 	if (mtag != NULL) {
1627 		/*
1628 		 * No need htons() here since vge(4) chip assumes
1629 		 * that tags are written in little endian and
1630 		 * we already use htole32() here.
1631 		 */
1632 		td_ctl |= VLAN_TAG_VALUE(mtag) | VGE_TDCTL_VTAG;
1633 	}
1634 	txd->td_ctl = htole32(td_ctl);
1635 	txd->td_sts = htole32(td_sts);
1636 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1637 
1638 	txd->td_sts = htole32(VGE_TDSTS_OWN | td_sts);
1639 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1640 
1641 	sc->sc_tx_free--;
1642 
1643 	return 0;
1644 }
1645 
1646 /*
1647  * Main transmit routine.
1648  */
1649 
1650 static void
1651 vge_start(struct ifnet *ifp)
1652 {
1653 	struct vge_softc *sc;
1654 	struct vge_txsoft *txs;
1655 	struct mbuf *m_head;
1656 	int idx, pidx, ofree, error;
1657 
1658 	sc = ifp->if_softc;
1659 
1660 	if (!sc->sc_link ||
1661 	    (ifp->if_flags & (IFF_RUNNING|IFF_OACTIVE)) != IFF_RUNNING) {
1662 		return;
1663 	}
1664 
1665 	m_head = NULL;
1666 	idx = sc->sc_tx_prodidx;
1667 	pidx = VGE_PREV_TXDESC(idx);
1668 	ofree = sc->sc_tx_free;
1669 
1670 	/*
1671 	 * Loop through the send queue, setting up transmit descriptors
1672 	 * until we drain the queue, or use up all available transmit
1673 	 * descriptors.
1674 	 */
1675 	for (;;) {
1676 		/* Grab a packet off the queue. */
1677 		IFQ_POLL(&ifp->if_snd, m_head);
1678 		if (m_head == NULL)
1679 			break;
1680 
1681 		if (sc->sc_tx_free == 0) {
1682 			/*
1683 			 * All slots used, stop for now.
1684 			 */
1685 			ifp->if_flags |= IFF_OACTIVE;
1686 			break;
1687 		}
1688 
1689 		txs = &sc->sc_txsoft[idx];
1690 		KASSERT(txs->txs_mbuf == NULL);
1691 
1692 		if ((error = vge_encap(sc, m_head, idx))) {
1693 			if (error == EFBIG) {
1694 				aprint_error_dev(&sc->sc_dev, "Tx packet consumes too many "
1695 				    "DMA segments, dropping...\n");
1696 				IFQ_DEQUEUE(&ifp->if_snd, m_head);
1697 				m_freem(m_head);
1698 				continue;
1699 			}
1700 
1701 			/*
1702 			 * Short on resources, just stop for now.
1703 			 */
1704 			if (error == ENOBUFS)
1705 				ifp->if_flags |= IFF_OACTIVE;
1706 			break;
1707 		}
1708 
1709 		IFQ_DEQUEUE(&ifp->if_snd, m_head);
1710 
1711 		/*
1712 		 * WE ARE NOW COMMITTED TO TRANSMITTING THE PACKET.
1713 		 */
1714 
1715 		sc->sc_txdescs[pidx].td_frag[0].tf_buflen |=
1716 		    htole16(VGE_TXDESC_Q);
1717 		VGE_TXFRAGSYNC(sc, pidx, 1,
1718 		    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1719 
1720 		if (txs->txs_mbuf != m_head) {
1721 			m_freem(m_head);
1722 			m_head = txs->txs_mbuf;
1723 		}
1724 
1725 		pidx = idx;
1726 		idx = VGE_NEXT_TXDESC(idx);
1727 
1728 		/*
1729 		 * If there's a BPF listener, bounce a copy of this frame
1730 		 * to him.
1731 		 */
1732 #if NBPFILTER > 0
1733 		if (ifp->if_bpf)
1734 			bpf_mtap(ifp->if_bpf, m_head);
1735 #endif
1736 	}
1737 
1738 	if (sc->sc_tx_free < ofree) {
1739 		/* TX packet queued */
1740 
1741 		sc->sc_tx_prodidx = idx;
1742 
1743 		/* Issue a transmit command. */
1744 		CSR_WRITE_2(sc, VGE_TXQCSRS, VGE_TXQCSR_WAK0);
1745 
1746 		/*
1747 		 * Use the countdown timer for interrupt moderation.
1748 		 * 'TX done' interrupts are disabled. Instead, we reset the
1749 		 * countdown timer, which will begin counting until it hits
1750 		 * the value in the SSTIMER register, and then trigger an
1751 		 * interrupt. Each time we set the TIMER0_ENABLE bit, the
1752 		 * the timer count is reloaded. Only when the transmitter
1753 		 * is idle will the timer hit 0 and an interrupt fire.
1754 		 */
1755 		CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_TIMER0_ENABLE);
1756 
1757 		/*
1758 		 * Set a timeout in case the chip goes out to lunch.
1759 		 */
1760 		ifp->if_timer = 5;
1761 	}
1762 }
1763 
1764 static int
1765 vge_init(struct ifnet *ifp)
1766 {
1767 	struct vge_softc *sc;
1768 	int i, rc = 0;
1769 
1770 	sc = ifp->if_softc;
1771 
1772 	/*
1773 	 * Cancel pending I/O and free all RX/TX buffers.
1774 	 */
1775 	vge_stop(sc);
1776 	vge_reset(sc);
1777 
1778 	/* Initialize the RX descriptors and mbufs. */
1779 	memset(sc->sc_rxdescs, 0, sizeof(sc->sc_rxdescs));
1780 	sc->sc_rx_consumed = 0;
1781 	for (i = 0; i < VGE_NRXDESC; i++) {
1782 		if (vge_newbuf(sc, i, NULL) == ENOBUFS) {
1783 			aprint_error_dev(&sc->sc_dev, "unable to allocate or map "
1784 			    "rx buffer\n");
1785 			return 1; /* XXX */
1786 		}
1787 	}
1788 	sc->sc_rx_prodidx = 0;
1789 	sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1790 
1791 	/* Initialize the  TX descriptors and mbufs. */
1792 	memset(sc->sc_txdescs, 0, sizeof(sc->sc_txdescs));
1793 	bus_dmamap_sync(sc->sc_dmat, sc->sc_cddmamap,
1794 	    VGE_CDTXOFF(0), sizeof(sc->sc_txdescs),
1795 	    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1796 	for (i = 0; i < VGE_NTXDESC; i++)
1797 		sc->sc_txsoft[i].txs_mbuf = NULL;
1798 
1799 	sc->sc_tx_prodidx = 0;
1800 	sc->sc_tx_considx = 0;
1801 	sc->sc_tx_free = VGE_NTXDESC;
1802 
1803 	/* Set our station address */
1804 	for (i = 0; i < ETHER_ADDR_LEN; i++)
1805 		CSR_WRITE_1(sc, VGE_PAR0 + i, sc->sc_eaddr[i]);
1806 
1807 	/*
1808 	 * Set receive FIFO threshold. Also allow transmission and
1809 	 * reception of VLAN tagged frames.
1810 	 */
1811 	CSR_CLRBIT_1(sc, VGE_RXCFG, VGE_RXCFG_FIFO_THR|VGE_RXCFG_VTAGOPT);
1812 	CSR_SETBIT_1(sc, VGE_RXCFG, VGE_RXFIFOTHR_128BYTES|VGE_VTAG_OPT2);
1813 
1814 	/* Set DMA burst length */
1815 	CSR_CLRBIT_1(sc, VGE_DMACFG0, VGE_DMACFG0_BURSTLEN);
1816 	CSR_SETBIT_1(sc, VGE_DMACFG0, VGE_DMABURST_128);
1817 
1818 	CSR_SETBIT_1(sc, VGE_TXCFG, VGE_TXCFG_ARB_PRIO|VGE_TXCFG_NONBLK);
1819 
1820 	/* Set collision backoff algorithm */
1821 	CSR_CLRBIT_1(sc, VGE_CHIPCFG1, VGE_CHIPCFG1_CRANDOM|
1822 	    VGE_CHIPCFG1_CAP|VGE_CHIPCFG1_MBA|VGE_CHIPCFG1_BAKOPT);
1823 	CSR_SETBIT_1(sc, VGE_CHIPCFG1, VGE_CHIPCFG1_OFSET);
1824 
1825 	/* Disable LPSEL field in priority resolution */
1826 	CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_LPSEL_DIS);
1827 
1828 	/*
1829 	 * Load the addresses of the DMA queues into the chip.
1830 	 * Note that we only use one transmit queue.
1831 	 */
1832 
1833 	CSR_WRITE_4(sc, VGE_TXDESC_ADDR_LO0, VGE_ADDR_LO(VGE_CDTXADDR(sc, 0)));
1834 	CSR_WRITE_2(sc, VGE_TXDESCNUM, VGE_NTXDESC - 1);
1835 
1836 	CSR_WRITE_4(sc, VGE_RXDESC_ADDR_LO, VGE_ADDR_LO(VGE_CDRXADDR(sc, 0)));
1837 	CSR_WRITE_2(sc, VGE_RXDESCNUM, VGE_NRXDESC - 1);
1838 	CSR_WRITE_2(sc, VGE_RXDESC_RESIDUECNT, VGE_NRXDESC);
1839 
1840 	/* Enable and wake up the RX descriptor queue */
1841 	CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_RUN);
1842 	CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_WAK);
1843 
1844 	/* Enable the TX descriptor queue */
1845 	CSR_WRITE_2(sc, VGE_TXQCSRS, VGE_TXQCSR_RUN0);
1846 
1847 	/* Set up the receive filter -- allow large frames for VLANs. */
1848 	CSR_WRITE_1(sc, VGE_RXCTL, VGE_RXCTL_RX_UCAST|VGE_RXCTL_RX_GIANT);
1849 
1850 	/* If we want promiscuous mode, set the allframes bit. */
1851 	if (ifp->if_flags & IFF_PROMISC) {
1852 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_PROMISC);
1853 	}
1854 
1855 	/* Set capture broadcast bit to capture broadcast frames. */
1856 	if (ifp->if_flags & IFF_BROADCAST) {
1857 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_BCAST);
1858 	}
1859 
1860 	/* Set multicast bit to capture multicast frames. */
1861 	if (ifp->if_flags & IFF_MULTICAST) {
1862 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_MCAST);
1863 	}
1864 
1865 	/* Init the cam filter. */
1866 	vge_cam_clear(sc);
1867 
1868 	/* Init the multicast filter. */
1869 	vge_setmulti(sc);
1870 
1871 	/* Enable flow control */
1872 
1873 	CSR_WRITE_1(sc, VGE_CRS2, 0x8B);
1874 
1875 	/* Enable jumbo frame reception (if desired) */
1876 
1877 	/* Start the MAC. */
1878 	CSR_WRITE_1(sc, VGE_CRC0, VGE_CR0_STOP);
1879 	CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_NOPOLL);
1880 	CSR_WRITE_1(sc, VGE_CRS0,
1881 	    VGE_CR0_TX_ENABLE|VGE_CR0_RX_ENABLE|VGE_CR0_START);
1882 
1883 	/*
1884 	 * Configure one-shot timer for microsecond
1885 	 * resulution and load it for 500 usecs.
1886 	 */
1887 	CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_TIMER0_RES);
1888 	CSR_WRITE_2(sc, VGE_SSTIMER, 400);
1889 
1890 	/*
1891 	 * Configure interrupt moderation for receive. Enable
1892 	 * the holdoff counter and load it, and set the RX
1893 	 * suppression count to the number of descriptors we
1894 	 * want to allow before triggering an interrupt.
1895 	 * The holdoff timer is in units of 20 usecs.
1896 	 */
1897 
1898 #ifdef notyet
1899 	CSR_WRITE_1(sc, VGE_INTCTL1, VGE_INTCTL_TXINTSUP_DISABLE);
1900 	/* Select the interrupt holdoff timer page. */
1901 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1902 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_INTHLDOFF);
1903 	CSR_WRITE_1(sc, VGE_INTHOLDOFF, 10); /* ~200 usecs */
1904 
1905 	/* Enable use of the holdoff timer. */
1906 	CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_HOLDOFF);
1907 	CSR_WRITE_1(sc, VGE_INTCTL1, VGE_INTCTL_SC_RELOAD);
1908 
1909 	/* Select the RX suppression threshold page. */
1910 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1911 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_RXSUPPTHR);
1912 	CSR_WRITE_1(sc, VGE_RXSUPPTHR, 64); /* interrupt after 64 packets */
1913 
1914 	/* Restore the page select bits. */
1915 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1916 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_MAR);
1917 #endif
1918 
1919 #ifdef DEVICE_POLLING
1920 	/*
1921 	 * Disable interrupts if we are polling.
1922 	 */
1923 	if (ifp->if_flags & IFF_POLLING) {
1924 		CSR_WRITE_4(sc, VGE_IMR, 0);
1925 		CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
1926 	} else	/* otherwise ... */
1927 #endif /* DEVICE_POLLING */
1928 	{
1929 	/*
1930 	 * Enable interrupts.
1931 	 */
1932 		CSR_WRITE_4(sc, VGE_IMR, VGE_INTRS);
1933 		CSR_WRITE_4(sc, VGE_ISR, 0);
1934 		CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_GMSK);
1935 	}
1936 
1937 	if ((rc = ether_mediachange(ifp)) != 0)
1938 		goto out;
1939 
1940 	ifp->if_flags |= IFF_RUNNING;
1941 	ifp->if_flags &= ~IFF_OACTIVE;
1942 
1943 	sc->sc_if_flags = 0;
1944 	sc->sc_link = 0;
1945 
1946 	callout_schedule(&sc->sc_timeout, hz);
1947 
1948 out:
1949 	return rc;
1950 }
1951 
1952 static void
1953 vge_miibus_statchg(struct device *self)
1954 {
1955 	struct vge_softc *sc;
1956 	struct mii_data *mii;
1957 	struct ifmedia_entry *ife;
1958 
1959 	sc = (void *)self;
1960 	mii = &sc->sc_mii;
1961 	ife = mii->mii_media.ifm_cur;
1962 	/*
1963 	 * If the user manually selects a media mode, we need to turn
1964 	 * on the forced MAC mode bit in the DIAGCTL register. If the
1965 	 * user happens to choose a full duplex mode, we also need to
1966 	 * set the 'force full duplex' bit. This applies only to
1967 	 * 10Mbps and 100Mbps speeds. In autoselect mode, forced MAC
1968 	 * mode is disabled, and in 1000baseT mode, full duplex is
1969 	 * always implied, so we turn on the forced mode bit but leave
1970 	 * the FDX bit cleared.
1971 	 */
1972 
1973 	switch (IFM_SUBTYPE(ife->ifm_media)) {
1974 	case IFM_AUTO:
1975 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
1976 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1977 		break;
1978 	case IFM_1000_T:
1979 		CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
1980 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1981 		break;
1982 	case IFM_100_TX:
1983 	case IFM_10_T:
1984 		CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
1985 		if ((ife->ifm_media & IFM_GMASK) == IFM_FDX) {
1986 			CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1987 		} else {
1988 			CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1989 		}
1990 		break;
1991 	default:
1992 		aprint_error_dev(&sc->sc_dev, "unknown media type: %x\n",
1993 		    IFM_SUBTYPE(ife->ifm_media));
1994 		break;
1995 	}
1996 }
1997 
1998 static int
1999 vge_ioctl(struct ifnet *ifp, u_long command, void *data)
2000 {
2001 	struct vge_softc *sc;
2002 	struct ifreq *ifr;
2003 	int s, error;
2004 
2005 	sc = ifp->if_softc;
2006 	ifr = (struct ifreq *)data;
2007 	error = 0;
2008 
2009 	s = splnet();
2010 
2011 	switch (command) {
2012 	case SIOCSIFMTU:
2013 		if (ifr->ifr_mtu > VGE_JUMBO_MTU)
2014 			error = EINVAL;
2015 		else if ((error = ifioctl_common(ifp, command, data)) == ENETRESET)
2016 			error = 0;
2017 		break;
2018 	case SIOCSIFFLAGS:
2019 		if (ifp->if_flags & IFF_UP) {
2020 			if (ifp->if_flags & IFF_RUNNING &&
2021 			    ifp->if_flags & IFF_PROMISC &&
2022 			    (sc->sc_if_flags & IFF_PROMISC) == 0) {
2023 				CSR_SETBIT_1(sc, VGE_RXCTL,
2024 				    VGE_RXCTL_RX_PROMISC);
2025 				vge_setmulti(sc);
2026 			} else if (ifp->if_flags & IFF_RUNNING &&
2027 			    (ifp->if_flags & IFF_PROMISC) == 0 &&
2028 			    sc->sc_if_flags & IFF_PROMISC) {
2029 				CSR_CLRBIT_1(sc, VGE_RXCTL,
2030 				    VGE_RXCTL_RX_PROMISC);
2031 				vge_setmulti(sc);
2032                         } else
2033 				vge_init(ifp);
2034 		} else {
2035 			if (ifp->if_flags & IFF_RUNNING)
2036 				vge_stop(sc);
2037 		}
2038 		sc->sc_if_flags = ifp->if_flags;
2039 		break;
2040 	default:
2041 		if ((error = ether_ioctl(ifp, command, data)) != ENETRESET)
2042 			break;
2043 
2044 		error = 0;
2045 
2046 		if (command != SIOCADDMULTI && command != SIOCDELMULTI)
2047 			;
2048 		else if (ifp->if_flags & IFF_RUNNING) {
2049 			/*
2050 			 * Multicast list has changed; set the hardware filter
2051 			 * accordingly.
2052 			 */
2053 			vge_setmulti(sc);
2054 		}
2055 		break;
2056 	}
2057 
2058 	splx(s);
2059 	return error;
2060 }
2061 
2062 static void
2063 vge_watchdog(struct ifnet *ifp)
2064 {
2065 	struct vge_softc *sc;
2066 	int s;
2067 
2068 	sc = ifp->if_softc;
2069 	s = splnet();
2070 	aprint_error_dev(&sc->sc_dev, "watchdog timeout\n");
2071 	ifp->if_oerrors++;
2072 
2073 	vge_txeof(sc);
2074 	vge_rxeof(sc);
2075 
2076 	vge_init(ifp);
2077 
2078 	splx(s);
2079 }
2080 
2081 /*
2082  * Stop the adapter and free any mbufs allocated to the
2083  * RX and TX lists.
2084  */
2085 static void
2086 vge_stop(struct vge_softc *sc)
2087 {
2088 	struct ifnet *ifp;
2089 	struct vge_txsoft *txs;
2090 	struct vge_rxsoft *rxs;
2091 	int i, s;
2092 
2093 	ifp = &sc->sc_ethercom.ec_if;
2094 
2095 	s = splnet();
2096 	ifp->if_timer = 0;
2097 
2098 	ifp->if_flags &= ~(IFF_RUNNING | IFF_OACTIVE);
2099 #ifdef DEVICE_POLLING
2100 	ether_poll_deregister(ifp);
2101 #endif /* DEVICE_POLLING */
2102 
2103 	CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
2104 	CSR_WRITE_1(sc, VGE_CRS0, VGE_CR0_STOP);
2105 	CSR_WRITE_4(sc, VGE_ISR, 0xFFFFFFFF);
2106 	CSR_WRITE_2(sc, VGE_TXQCSRC, 0xFFFF);
2107 	CSR_WRITE_1(sc, VGE_RXQCSRC, 0xFF);
2108 	CSR_WRITE_4(sc, VGE_RXDESC_ADDR_LO, 0);
2109 
2110 	if (sc->sc_rx_mhead != NULL) {
2111 		m_freem(sc->sc_rx_mhead);
2112 		sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
2113 	}
2114 
2115 	/* Free the TX list buffers. */
2116 
2117 	for (i = 0; i < VGE_NTXDESC; i++) {
2118 		txs = &sc->sc_txsoft[i];
2119 		if (txs->txs_mbuf != NULL) {
2120 			bus_dmamap_unload(sc->sc_dmat, txs->txs_dmamap);
2121 			m_freem(txs->txs_mbuf);
2122 			txs->txs_mbuf = NULL;
2123 		}
2124 	}
2125 
2126 	/* Free the RX list buffers. */
2127 
2128 	for (i = 0; i < VGE_NRXDESC; i++) {
2129 		rxs = &sc->sc_rxsoft[i];
2130 		if (rxs->rxs_mbuf != NULL) {
2131 			bus_dmamap_unload(sc->sc_dmat, rxs->rxs_dmamap);
2132 			m_freem(rxs->rxs_mbuf);
2133 			rxs->rxs_mbuf = NULL;
2134 		}
2135 	}
2136 
2137 	splx(s);
2138 }
2139 
2140 #if VGE_POWER_MANAGEMENT
2141 /*
2142  * Device suspend routine.  Stop the interface and save some PCI
2143  * settings in case the BIOS doesn't restore them properly on
2144  * resume.
2145  */
2146 static int
2147 vge_suspend(struct device *dev)
2148 {
2149 	struct vge_softc *sc;
2150 	int i;
2151 
2152 	sc = device_get_softc(dev);
2153 
2154 	vge_stop(sc);
2155 
2156         for (i = 0; i < 5; i++)
2157 		sc->sc_saved_maps[i] =
2158 		    pci_read_config(dev, PCIR_MAPS + i * 4, 4);
2159 	sc->sc_saved_biosaddr = pci_read_config(dev, PCIR_BIOS, 4);
2160 	sc->sc_saved_intline = pci_read_config(dev, PCIR_INTLINE, 1);
2161 	sc->sc_saved_cachelnsz = pci_read_config(dev, PCIR_CACHELNSZ, 1);
2162 	sc->sc_saved_lattimer = pci_read_config(dev, PCIR_LATTIMER, 1);
2163 
2164 	sc->suspended = 1;
2165 
2166 	return 0;
2167 }
2168 
2169 /*
2170  * Device resume routine.  Restore some PCI settings in case the BIOS
2171  * doesn't, re-enable busmastering, and restart the interface if
2172  * appropriate.
2173  */
2174 static int
2175 vge_resume(struct device *dev)
2176 {
2177 	struct vge_softc *sc;
2178 	struct ifnet *ifp;
2179 	int i;
2180 
2181 	sc = (void *)dev;
2182 	ifp = &sc->sc_ethercom.ec_if;
2183 
2184         /* better way to do this? */
2185 	for (i = 0; i < 5; i++)
2186 		pci_write_config(dev, PCIR_MAPS + i * 4,
2187 		    sc->sc_saved_maps[i], 4);
2188 	pci_write_config(dev, PCIR_BIOS, sc->sc_saved_biosaddr, 4);
2189 	pci_write_config(dev, PCIR_INTLINE, sc->sc_saved_intline, 1);
2190 	pci_write_config(dev, PCIR_CACHELNSZ, sc->sc_saved_cachelnsz, 1);
2191 	pci_write_config(dev, PCIR_LATTIMER, sc->sc_saved_lattimer, 1);
2192 
2193 	/* reenable busmastering */
2194 	pci_enable_busmaster(dev);
2195 	pci_enable_io(dev, SYS_RES_MEMORY);
2196 
2197 	/* reinitialize interface if necessary */
2198 	if (ifp->if_flags & IFF_UP)
2199 		vge_init(sc);
2200 
2201 	sc->suspended = 0;
2202 
2203 	return 0;
2204 }
2205 #endif
2206 
2207 /*
2208  * Stop all chip I/O so that the kernel's probe routines don't
2209  * get confused by errant DMAs when rebooting.
2210  */
2211 static void
2212 vge_shutdown(void *arg)
2213 {
2214 	struct vge_softc *sc;
2215 
2216 	sc = arg;
2217 	vge_stop(sc);
2218 }
2219