xref: /netbsd-src/sys/dev/pci/if_vge.c (revision 3816d47b2c42fcd6e549e3407f842a5b1a1d23ad)
1 /* $NetBSD: if_vge.c,v 1.49 2009/09/07 12:44:29 tsutsui 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 1.49 2009/09/07 12:44:29 tsutsui 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 	device_t		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(device_t, cfdata_t, void *);
299 static void vge_attach(device_t, device_t, 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(device_t);
319 static int vge_resume(device_t);
320 #endif
321 static bool vge_shutdown(device_t, int);
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(device_t, int, int);
328 static void vge_miibus_writereg(device_t, int, int, int);
329 static void vge_miibus_statchg(device_t);
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_NEW(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 		printf("%s: EEPROM read timed out\n", device_xname(sc->sc_dev));
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 		printf("%s: failed to idle MII autopoll\n",
478 		    device_xname(sc->sc_dev));
479 	}
480 }
481 
482 static void
483 vge_miipoll_start(struct vge_softc *sc)
484 {
485 	int i;
486 
487 	/* First, make sure we're idle. */
488 
489 	CSR_WRITE_1(sc, VGE_MIICMD, 0);
490 	CSR_WRITE_1(sc, VGE_MIIADDR, VGE_MIIADDR_SWMPL);
491 
492 	for (i = 0; i < VGE_TIMEOUT; i++) {
493 		DELAY(1);
494 		if (CSR_READ_1(sc, VGE_MIISTS) & VGE_MIISTS_IIDL)
495 			break;
496 	}
497 
498 	if (i == VGE_TIMEOUT) {
499 		printf("%s: failed to idle MII autopoll\n",
500 		    device_xname(sc->sc_dev));
501 		return;
502 	}
503 
504 	/* Now enable auto poll mode. */
505 
506 	CSR_WRITE_1(sc, VGE_MIICMD, VGE_MIICMD_MAUTO);
507 
508 	/* And make sure it started. */
509 
510 	for (i = 0; i < VGE_TIMEOUT; i++) {
511 		DELAY(1);
512 		if ((CSR_READ_1(sc, VGE_MIISTS) & VGE_MIISTS_IIDL) == 0)
513 			break;
514 	}
515 
516 	if (i == VGE_TIMEOUT) {
517 		printf("%s: failed to start MII autopoll\n",
518 		    device_xname(sc->sc_dev));
519 	}
520 }
521 
522 static int
523 vge_miibus_readreg(device_t dev, int phy, int reg)
524 {
525 	struct vge_softc *sc;
526 	int i, s;
527 	uint16_t rval;
528 
529 	sc = device_private(dev);
530 	rval = 0;
531 	if (phy != (CSR_READ_1(sc, VGE_MIICFG) & 0x1F))
532 		return 0;
533 
534 	s = splnet();
535 	vge_miipoll_stop(sc);
536 
537 	/* Specify the register we want to read. */
538 	CSR_WRITE_1(sc, VGE_MIIADDR, reg);
539 
540 	/* Issue read command. */
541 	CSR_SETBIT_1(sc, VGE_MIICMD, VGE_MIICMD_RCMD);
542 
543 	/* Wait for the read command bit to self-clear. */
544 	for (i = 0; i < VGE_TIMEOUT; i++) {
545 		DELAY(1);
546 		if ((CSR_READ_1(sc, VGE_MIICMD) & VGE_MIICMD_RCMD) == 0)
547 			break;
548 	}
549 
550 	if (i == VGE_TIMEOUT)
551 		printf("%s: MII read timed out\n", device_xname(sc->sc_dev));
552 	else
553 		rval = CSR_READ_2(sc, VGE_MIIDATA);
554 
555 	vge_miipoll_start(sc);
556 	splx(s);
557 
558 	return rval;
559 }
560 
561 static void
562 vge_miibus_writereg(device_t dev, int phy, int reg, int data)
563 {
564 	struct vge_softc *sc;
565 	int i, s;
566 
567 	sc = device_private(dev);
568 	if (phy != (CSR_READ_1(sc, VGE_MIICFG) & 0x1F))
569 		return;
570 
571 	s = splnet();
572 	vge_miipoll_stop(sc);
573 
574 	/* Specify the register we want to write. */
575 	CSR_WRITE_1(sc, VGE_MIIADDR, reg);
576 
577 	/* Specify the data we want to write. */
578 	CSR_WRITE_2(sc, VGE_MIIDATA, data);
579 
580 	/* Issue write command. */
581 	CSR_SETBIT_1(sc, VGE_MIICMD, VGE_MIICMD_WCMD);
582 
583 	/* Wait for the write command bit to self-clear. */
584 	for (i = 0; i < VGE_TIMEOUT; i++) {
585 		DELAY(1);
586 		if ((CSR_READ_1(sc, VGE_MIICMD) & VGE_MIICMD_WCMD) == 0)
587 			break;
588 	}
589 
590 	if (i == VGE_TIMEOUT) {
591 		printf("%s: MII write timed out\n", device_xname(sc->sc_dev));
592 	}
593 
594 	vge_miipoll_start(sc);
595 	splx(s);
596 }
597 
598 static void
599 vge_cam_clear(struct vge_softc *sc)
600 {
601 	int i;
602 
603 	/*
604 	 * Turn off all the mask bits. This tells the chip
605 	 * that none of the entries in the CAM filter are valid.
606 	 * desired entries will be enabled as we fill the filter in.
607 	 */
608 
609 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
610 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_CAMMASK);
611 	CSR_WRITE_1(sc, VGE_CAMADDR, VGE_CAMADDR_ENABLE);
612 	for (i = 0; i < 8; i++)
613 		CSR_WRITE_1(sc, VGE_CAM0 + i, 0);
614 
615 	/* Clear the VLAN filter too. */
616 
617 	CSR_WRITE_1(sc, VGE_CAMADDR, VGE_CAMADDR_ENABLE|VGE_CAMADDR_AVSEL|0);
618 	for (i = 0; i < 8; i++)
619 		CSR_WRITE_1(sc, VGE_CAM0 + i, 0);
620 
621 	CSR_WRITE_1(sc, VGE_CAMADDR, 0);
622 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
623 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_MAR);
624 
625 	sc->sc_camidx = 0;
626 }
627 
628 static int
629 vge_cam_set(struct vge_softc *sc, uint8_t *addr)
630 {
631 	int i, error;
632 
633 	error = 0;
634 
635 	if (sc->sc_camidx == VGE_CAM_MAXADDRS)
636 		return ENOSPC;
637 
638 	/* Select the CAM data page. */
639 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
640 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_CAMDATA);
641 
642 	/* Set the filter entry we want to update and enable writing. */
643 	CSR_WRITE_1(sc, VGE_CAMADDR, VGE_CAMADDR_ENABLE | sc->sc_camidx);
644 
645 	/* Write the address to the CAM registers */
646 	for (i = 0; i < ETHER_ADDR_LEN; i++)
647 		CSR_WRITE_1(sc, VGE_CAM0 + i, addr[i]);
648 
649 	/* Issue a write command. */
650 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_WRITE);
651 
652 	/* Wake for it to clear. */
653 	for (i = 0; i < VGE_TIMEOUT; i++) {
654 		DELAY(1);
655 		if ((CSR_READ_1(sc, VGE_CAMCTL) & VGE_CAMCTL_WRITE) == 0)
656 			break;
657 	}
658 
659 	if (i == VGE_TIMEOUT) {
660 		printf("%s: setting CAM filter failed\n",
661 		    device_xname(sc->sc_dev));
662 		error = EIO;
663 		goto fail;
664 	}
665 
666 	/* Select the CAM mask page. */
667 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
668 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_CAMMASK);
669 
670 	/* Set the mask bit that enables this filter. */
671 	CSR_SETBIT_1(sc, VGE_CAM0 + (sc->sc_camidx / 8),
672 	    1 << (sc->sc_camidx & 7));
673 
674 	sc->sc_camidx++;
675 
676  fail:
677 	/* Turn off access to CAM. */
678 	CSR_WRITE_1(sc, VGE_CAMADDR, 0);
679 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
680 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_MAR);
681 
682 	return error;
683 }
684 
685 /*
686  * Program the multicast filter. We use the 64-entry CAM filter
687  * for perfect filtering. If there's more than 64 multicast addresses,
688  * we use the hash filter instead.
689  */
690 static void
691 vge_setmulti(struct vge_softc *sc)
692 {
693 	struct ifnet *ifp;
694 	int error;
695 	uint32_t h, hashes[2] = { 0, 0 };
696 	struct ether_multi *enm;
697 	struct ether_multistep step;
698 
699 	error = 0;
700 	ifp = &sc->sc_ethercom.ec_if;
701 
702 	/* First, zot all the multicast entries. */
703 	vge_cam_clear(sc);
704 	CSR_WRITE_4(sc, VGE_MAR0, 0);
705 	CSR_WRITE_4(sc, VGE_MAR1, 0);
706 	ifp->if_flags &= ~IFF_ALLMULTI;
707 
708 	/*
709 	 * If the user wants allmulti or promisc mode, enable reception
710 	 * of all multicast frames.
711 	 */
712 	if (ifp->if_flags & IFF_PROMISC) {
713  allmulti:
714 		CSR_WRITE_4(sc, VGE_MAR0, 0xFFFFFFFF);
715 		CSR_WRITE_4(sc, VGE_MAR1, 0xFFFFFFFF);
716 		ifp->if_flags |= IFF_ALLMULTI;
717 		return;
718 	}
719 
720 	/* Now program new ones */
721 	ETHER_FIRST_MULTI(step, &sc->sc_ethercom, enm);
722 	while (enm != NULL) {
723 		/*
724 		 * If multicast range, fall back to ALLMULTI.
725 		 */
726 		if (memcmp(enm->enm_addrlo, enm->enm_addrhi,
727 				ETHER_ADDR_LEN) != 0)
728 			goto allmulti;
729 
730 		error = vge_cam_set(sc, enm->enm_addrlo);
731 		if (error)
732 			break;
733 
734 		ETHER_NEXT_MULTI(step, enm);
735 	}
736 
737 	/* If there were too many addresses, use the hash filter. */
738 	if (error) {
739 		vge_cam_clear(sc);
740 
741 		ETHER_FIRST_MULTI(step, &sc->sc_ethercom, enm);
742 		while (enm != NULL) {
743 			/*
744 			 * If multicast range, fall back to ALLMULTI.
745 			 */
746 			if (memcmp(enm->enm_addrlo, enm->enm_addrhi,
747 					ETHER_ADDR_LEN) != 0)
748 				goto allmulti;
749 
750 			h = ether_crc32_be(enm->enm_addrlo,
751 			    ETHER_ADDR_LEN) >> 26;
752 			hashes[h >> 5] |= 1 << (h & 0x1f);
753 
754 			ETHER_NEXT_MULTI(step, enm);
755 		}
756 
757 		CSR_WRITE_4(sc, VGE_MAR0, hashes[0]);
758 		CSR_WRITE_4(sc, VGE_MAR1, hashes[1]);
759 	}
760 }
761 
762 static void
763 vge_reset(struct vge_softc *sc)
764 {
765 	int i;
766 
767 	CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_SOFTRESET);
768 
769 	for (i = 0; i < VGE_TIMEOUT; i++) {
770 		DELAY(5);
771 		if ((CSR_READ_1(sc, VGE_CRS1) & VGE_CR1_SOFTRESET) == 0)
772 			break;
773 	}
774 
775 	if (i == VGE_TIMEOUT) {
776 		printf("%s: soft reset timed out", device_xname(sc->sc_dev));
777 		CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_STOP_FORCE);
778 		DELAY(2000);
779 	}
780 
781 	DELAY(5000);
782 
783 	CSR_SETBIT_1(sc, VGE_EECSR, VGE_EECSR_RELOAD);
784 
785 	for (i = 0; i < VGE_TIMEOUT; i++) {
786 		DELAY(5);
787 		if ((CSR_READ_1(sc, VGE_EECSR) & VGE_EECSR_RELOAD) == 0)
788 			break;
789 	}
790 
791 	if (i == VGE_TIMEOUT) {
792 		printf("%s: EEPROM reload timed out\n",
793 		    device_xname(sc->sc_dev));
794 		return;
795 	}
796 
797 	/*
798 	 * On some machine, the first read data from EEPROM could be
799 	 * messed up, so read one dummy data here to avoid the mess.
800 	 */
801 	(void)vge_read_eeprom(sc, 0);
802 
803 	CSR_CLRBIT_1(sc, VGE_CHIPCFG0, VGE_CHIPCFG0_PACPI);
804 }
805 
806 /*
807  * Probe for a VIA gigabit chip. Check the PCI vendor and device
808  * IDs against our list and return a device name if we find a match.
809  */
810 static int
811 vge_match(device_t parent, cfdata_t match, void *aux)
812 {
813 	struct pci_attach_args *pa = aux;
814 
815 	if (PCI_VENDOR(pa->pa_id) == PCI_VENDOR_VIATECH
816 	    && PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_VIATECH_VT612X)
817 		return 1;
818 
819 	return 0;
820 }
821 
822 static int
823 vge_allocmem(struct vge_softc *sc)
824 {
825 	int error;
826 	int nseg;
827 	int i;
828 	bus_dma_segment_t seg;
829 
830 	/*
831 	 * Allocate memory for control data.
832 	 */
833 
834 	error = bus_dmamem_alloc(sc->sc_dmat, sizeof(struct vge_control_data),
835 	     VGE_RING_ALIGN, 0, &seg, 1, &nseg, BUS_DMA_NOWAIT);
836 	if (error) {
837 		aprint_error_dev(sc->sc_dev,
838 		    "could not allocate control data dma memory\n");
839 		goto fail_1;
840 	}
841 
842 	/* Map the memory to kernel VA space */
843 
844 	error = bus_dmamem_map(sc->sc_dmat, &seg, nseg,
845 	    sizeof(struct vge_control_data), (void **)&sc->sc_control_data,
846 	    BUS_DMA_NOWAIT);
847 	if (error) {
848 		aprint_error_dev(sc->sc_dev,
849 		    "could not map control data dma memory\n");
850 		goto fail_2;
851 	}
852 	memset(sc->sc_control_data, 0, sizeof(struct vge_control_data));
853 
854 	/*
855 	 * Create map for control data.
856 	 */
857 	error = bus_dmamap_create(sc->sc_dmat,
858 	    sizeof(struct vge_control_data), 1,
859 	    sizeof(struct vge_control_data), 0, BUS_DMA_NOWAIT,
860 	    &sc->sc_cddmamap);
861 	if (error) {
862 		aprint_error_dev(sc->sc_dev,
863 		    "could not create control data dmamap\n");
864 		goto fail_3;
865 	}
866 
867 	/* Load the map for the control data. */
868 	error = bus_dmamap_load(sc->sc_dmat, sc->sc_cddmamap,
869 	    sc->sc_control_data, sizeof(struct vge_control_data), NULL,
870 	    BUS_DMA_NOWAIT);
871 	if (error) {
872 		aprint_error_dev(sc->sc_dev,
873 		    "could not load control data dma memory\n");
874 		goto fail_4;
875 	}
876 
877 	/* Create DMA maps for TX buffers */
878 
879 	for (i = 0; i < VGE_NTXDESC; i++) {
880 		error = bus_dmamap_create(sc->sc_dmat, VGE_TX_MAXLEN,
881 		    VGE_TX_FRAGS, VGE_TX_MAXLEN, 0, BUS_DMA_NOWAIT,
882 		    &sc->sc_txsoft[i].txs_dmamap);
883 		if (error) {
884 			aprint_error_dev(sc->sc_dev,
885 			    "can't create DMA map for TX descs\n");
886 			goto fail_5;
887 		}
888 	}
889 
890 	/* Create DMA maps for RX buffers */
891 
892 	for (i = 0; i < VGE_NRXDESC; i++) {
893 		error = bus_dmamap_create(sc->sc_dmat, MCLBYTES,
894 		    1, MCLBYTES, 0, BUS_DMA_NOWAIT,
895 		    &sc->sc_rxsoft[i].rxs_dmamap);
896 		if (error) {
897 			aprint_error_dev(sc->sc_dev,
898 			    "can't create DMA map for RX descs\n");
899 			goto fail_6;
900 		}
901 		sc->sc_rxsoft[i].rxs_mbuf = NULL;
902 	}
903 
904 	return 0;
905 
906  fail_6:
907 	for (i = 0; i < VGE_NRXDESC; i++) {
908 		if (sc->sc_rxsoft[i].rxs_dmamap != NULL)
909 			bus_dmamap_destroy(sc->sc_dmat,
910 			    sc->sc_rxsoft[i].rxs_dmamap);
911 	}
912  fail_5:
913 	for (i = 0; i < VGE_NTXDESC; i++) {
914 		if (sc->sc_txsoft[i].txs_dmamap != NULL)
915 			bus_dmamap_destroy(sc->sc_dmat,
916 			    sc->sc_txsoft[i].txs_dmamap);
917 	}
918 	bus_dmamap_unload(sc->sc_dmat, sc->sc_cddmamap);
919  fail_4:
920 	bus_dmamap_destroy(sc->sc_dmat, sc->sc_cddmamap);
921  fail_3:
922 	bus_dmamem_unmap(sc->sc_dmat, (void *)sc->sc_control_data,
923 	    sizeof(struct vge_control_data));
924  fail_2:
925 	bus_dmamem_free(sc->sc_dmat, &seg, nseg);
926  fail_1:
927 	return ENOMEM;
928 }
929 
930 /*
931  * Attach the interface. Allocate softc structures, do ifmedia
932  * setup and ethernet/BPF attach.
933  */
934 static void
935 vge_attach(device_t parent, device_t self, void *aux)
936 {
937 	uint8_t	*eaddr;
938 	struct vge_softc *sc = device_private(self);
939 	struct ifnet *ifp;
940 	struct pci_attach_args *pa = aux;
941 	pci_chipset_tag_t pc = pa->pa_pc;
942 	const char *intrstr;
943 	pci_intr_handle_t ih;
944 	uint16_t val;
945 
946 	sc->sc_dev = self;
947 
948 	aprint_normal(": VIA VT612X Gigabit Ethernet (rev. %#x)\n",
949 	    PCI_REVISION(pa->pa_class));
950 
951 	/* Make sure bus-mastering is enabled */
952         pci_conf_write(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG,
953 	    pci_conf_read(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG) |
954 	    PCI_COMMAND_MASTER_ENABLE);
955 
956 	/*
957 	 * Map control/status registers.
958 	 */
959 	if (pci_mapreg_map(pa, VGE_PCI_LOMEM, PCI_MAPREG_TYPE_MEM, 0,
960 	    &sc->sc_bst, &sc->sc_bsh, NULL, NULL) != 0) {
961 		aprint_error_dev(self, "couldn't map memory\n");
962 		return;
963 	}
964 
965         /*
966          * Map and establish our interrupt.
967          */
968 	if (pci_intr_map(pa, &ih)) {
969 		aprint_error_dev(self, "unable to map interrupt\n");
970 		return;
971 	}
972 	intrstr = pci_intr_string(pc, ih);
973 	sc->sc_intrhand = pci_intr_establish(pc, ih, IPL_NET, vge_intr, sc);
974 	if (sc->sc_intrhand == NULL) {
975 		aprint_error_dev(self, "unable to establish interrupt");
976 		if (intrstr != NULL)
977 			aprint_error(" at %s", intrstr);
978 		aprint_error("\n");
979 		return;
980 	}
981 	aprint_normal_dev(self, "interrupting at %s\n", intrstr);
982 
983 	/* Reset the adapter. */
984 	vge_reset(sc);
985 
986 	/*
987 	 * Get station address from the EEPROM.
988 	 */
989 	eaddr = sc->sc_eaddr;
990 	val = vge_read_eeprom(sc, VGE_EE_EADDR + 0);
991 	eaddr[0] = val & 0xff;
992 	eaddr[1] = val >> 8;
993 	val = vge_read_eeprom(sc, VGE_EE_EADDR + 1);
994 	eaddr[2] = val & 0xff;
995 	eaddr[3] = val >> 8;
996 	val = vge_read_eeprom(sc, VGE_EE_EADDR + 2);
997 	eaddr[4] = val & 0xff;
998 	eaddr[5] = val >> 8;
999 
1000 	aprint_normal_dev(self, "Ethernet address: %s\n",
1001 	    ether_sprintf(eaddr));
1002 
1003 	/*
1004 	 * Use the 32bit tag. Hardware supports 48bit physical addresses,
1005 	 * but we don't use that for now.
1006 	 */
1007 	sc->sc_dmat = pa->pa_dmat;
1008 
1009 	if (vge_allocmem(sc) != 0)
1010 		return;
1011 
1012 	ifp = &sc->sc_ethercom.ec_if;
1013 	ifp->if_softc = sc;
1014 	strlcpy(ifp->if_xname, device_xname(self), IFNAMSIZ);
1015 	ifp->if_mtu = ETHERMTU;
1016 	ifp->if_baudrate = IF_Gbps(1);
1017 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
1018 	ifp->if_ioctl = vge_ioctl;
1019 	ifp->if_start = vge_start;
1020 	ifp->if_init = vge_init;
1021 	ifp->if_stop = vge_stop;
1022 
1023 	/*
1024 	 * We can support 802.1Q VLAN-sized frames and jumbo
1025 	 * Ethernet frames.
1026 	 */
1027 	sc->sc_ethercom.ec_capabilities |=
1028 	    ETHERCAP_VLAN_MTU | ETHERCAP_JUMBO_MTU |
1029 	    ETHERCAP_VLAN_HWTAGGING;
1030 
1031 	/*
1032 	 * We can do IPv4/TCPv4/UDPv4 checksums in hardware.
1033 	 */
1034 	ifp->if_capabilities |=
1035 	    IFCAP_CSUM_IPv4_Tx | IFCAP_CSUM_IPv4_Rx |
1036 	    IFCAP_CSUM_TCPv4_Tx | IFCAP_CSUM_TCPv4_Rx |
1037 	    IFCAP_CSUM_UDPv4_Tx | IFCAP_CSUM_UDPv4_Rx;
1038 
1039 #ifdef DEVICE_POLLING
1040 #ifdef IFCAP_POLLING
1041 	ifp->if_capabilities |= IFCAP_POLLING;
1042 #endif
1043 #endif
1044 	ifp->if_watchdog = vge_watchdog;
1045 	IFQ_SET_MAXLEN(&ifp->if_snd, max(VGE_IFQ_MAXLEN, IFQ_MAXLEN));
1046 	IFQ_SET_READY(&ifp->if_snd);
1047 
1048 	/*
1049 	 * Initialize our media structures and probe the MII.
1050 	 */
1051 	sc->sc_mii.mii_ifp = ifp;
1052 	sc->sc_mii.mii_readreg = vge_miibus_readreg;
1053 	sc->sc_mii.mii_writereg = vge_miibus_writereg;
1054 	sc->sc_mii.mii_statchg = vge_miibus_statchg;
1055 
1056 	sc->sc_ethercom.ec_mii = &sc->sc_mii;
1057 	ifmedia_init(&sc->sc_mii.mii_media, 0, ether_mediachange,
1058 	    ether_mediastatus);
1059 	mii_attach(self, &sc->sc_mii, 0xffffffff, MII_PHY_ANY,
1060 	    MII_OFFSET_ANY, MIIF_DOPAUSE);
1061 	if (LIST_FIRST(&sc->sc_mii.mii_phys) == NULL) {
1062 		ifmedia_add(&sc->sc_mii.mii_media, IFM_ETHER|IFM_NONE, 0, NULL);
1063 		ifmedia_set(&sc->sc_mii.mii_media, IFM_ETHER|IFM_NONE);
1064 	} else
1065 		ifmedia_set(&sc->sc_mii.mii_media, IFM_ETHER|IFM_AUTO);
1066 
1067 	/*
1068 	 * Attach the interface.
1069 	 */
1070 	if_attach(ifp);
1071 	ether_ifattach(ifp, eaddr);
1072 	ether_set_ifflags_cb(&sc->sc_ethercom, vge_ifflags_cb);
1073 
1074 	callout_init(&sc->sc_timeout, 0);
1075 	callout_setfunc(&sc->sc_timeout, vge_tick, sc);
1076 
1077 	/*
1078 	 * Make sure the interface is shutdown during reboot.
1079 	 */
1080 	if (pmf_device_register1(self, NULL, NULL, vge_shutdown))
1081 		pmf_class_network_register(self, ifp);
1082 	else
1083 		aprint_error_dev(self, "couldn't establish power handler\n");
1084 }
1085 
1086 static int
1087 vge_newbuf(struct vge_softc *sc, int idx, struct mbuf *m)
1088 {
1089 	struct mbuf *m_new;
1090 	struct vge_rxdesc *rxd;
1091 	struct vge_rxsoft *rxs;
1092 	bus_dmamap_t map;
1093 	int i;
1094 #ifdef DIAGNOSTIC
1095 	uint32_t rd_sts;
1096 #endif
1097 
1098 	m_new = NULL;
1099 	if (m == NULL) {
1100 		MGETHDR(m_new, M_DONTWAIT, MT_DATA);
1101 		if (m_new == NULL)
1102 			return ENOBUFS;
1103 
1104 		MCLGET(m_new, M_DONTWAIT);
1105 		if ((m_new->m_flags & M_EXT) == 0) {
1106 			m_freem(m_new);
1107 			return ENOBUFS;
1108 		}
1109 
1110 		m = m_new;
1111 	} else
1112 		m->m_data = m->m_ext.ext_buf;
1113 
1114 
1115 	/*
1116 	 * This is part of an evil trick to deal with non-x86 platforms.
1117 	 * The VIA chip requires RX buffers to be aligned on 32-bit
1118 	 * boundaries, but that will hose non-x86 machines. To get around
1119 	 * this, we leave some empty space at the start of each buffer
1120 	 * and for non-x86 hosts, we copy the buffer back two bytes
1121 	 * to achieve word alignment. This is slightly more efficient
1122 	 * than allocating a new buffer, copying the contents, and
1123 	 * discarding the old buffer.
1124 	 */
1125 	m->m_len = m->m_pkthdr.len = VGE_RX_BUFSIZE;
1126 #ifndef __NO_STRICT_ALIGNMENT
1127 	m->m_data += VGE_RX_PAD;
1128 #endif
1129 	rxs = &sc->sc_rxsoft[idx];
1130 	map = rxs->rxs_dmamap;
1131 
1132 	if (bus_dmamap_load_mbuf(sc->sc_dmat, map, m, BUS_DMA_NOWAIT) != 0)
1133 		goto out;
1134 
1135 	rxd = &sc->sc_rxdescs[idx];
1136 
1137 #ifdef DIAGNOSTIC
1138 	/* If this descriptor is still owned by the chip, bail. */
1139 	VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1140 	rd_sts = le32toh(rxd->rd_sts);
1141 	VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1142 	if (rd_sts & VGE_RDSTS_OWN) {
1143 		panic("%s: tried to map busy RX descriptor",
1144 		    device_xname(sc->sc_dev));
1145 	}
1146 #endif
1147 
1148 	rxs->rxs_mbuf = m;
1149 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1150 	    BUS_DMASYNC_PREREAD);
1151 
1152 	rxd->rd_buflen =
1153 	    htole16(VGE_BUFLEN(map->dm_segs[0].ds_len) | VGE_RXDESC_I);
1154 	vge_set_rxaddr(rxd, map->dm_segs[0].ds_addr);
1155 	rxd->rd_sts = 0;
1156 	rxd->rd_ctl = 0;
1157 	VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1158 
1159 	/*
1160 	 * Note: the manual fails to document the fact that for
1161 	 * proper opration, the driver needs to replentish the RX
1162 	 * DMA ring 4 descriptors at a time (rather than one at a
1163 	 * time, like most chips). We can allocate the new buffers
1164 	 * but we should not set the OWN bits until we're ready
1165 	 * to hand back 4 of them in one shot.
1166 	 */
1167 
1168 #define VGE_RXCHUNK 4
1169 	sc->sc_rx_consumed++;
1170 	if (sc->sc_rx_consumed == VGE_RXCHUNK) {
1171 		for (i = idx; i != idx - VGE_RXCHUNK; i--) {
1172 			KASSERT(i >= 0);
1173 			sc->sc_rxdescs[i].rd_sts |= htole32(VGE_RDSTS_OWN);
1174 			VGE_RXDESCSYNC(sc, i,
1175 			    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1176 		}
1177 		sc->sc_rx_consumed = 0;
1178 	}
1179 
1180 	return 0;
1181  out:
1182 	if (m_new != NULL)
1183 		m_freem(m_new);
1184 	return ENOMEM;
1185 }
1186 
1187 #ifndef __NO_STRICT_ALIGNMENT
1188 static inline void
1189 vge_fixup_rx(struct mbuf *m)
1190 {
1191 	int i;
1192 	uint16_t *src, *dst;
1193 
1194 	src = mtod(m, uint16_t *);
1195 	dst = src - 1;
1196 
1197 	for (i = 0; i < (m->m_len / sizeof(uint16_t) + 1); i++)
1198 		*dst++ = *src++;
1199 
1200 	m->m_data -= ETHER_ALIGN;
1201 }
1202 #endif
1203 
1204 /*
1205  * RX handler. We support the reception of jumbo frames that have
1206  * been fragmented across multiple 2K mbuf cluster buffers.
1207  */
1208 static void
1209 vge_rxeof(struct vge_softc *sc)
1210 {
1211 	struct mbuf *m;
1212 	struct ifnet *ifp;
1213 	int idx, total_len, lim;
1214 	struct vge_rxdesc *cur_rxd;
1215 	struct vge_rxsoft *rxs;
1216 	uint32_t rxstat, rxctl;
1217 
1218 	ifp = &sc->sc_ethercom.ec_if;
1219 	lim = 0;
1220 
1221 	/* Invalidate the descriptor memory */
1222 
1223 	for (idx = sc->sc_rx_prodidx;; idx = VGE_NEXT_RXDESC(idx)) {
1224 		cur_rxd = &sc->sc_rxdescs[idx];
1225 
1226 		VGE_RXDESCSYNC(sc, idx,
1227 		    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1228 		rxstat = le32toh(cur_rxd->rd_sts);
1229 		if ((rxstat & VGE_RDSTS_OWN) != 0) {
1230 			VGE_RXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1231 			break;
1232 		}
1233 
1234 		rxctl = le32toh(cur_rxd->rd_ctl);
1235 		rxs = &sc->sc_rxsoft[idx];
1236 		m = rxs->rxs_mbuf;
1237 		total_len = (rxstat & VGE_RDSTS_BUFSIZ) >> 16;
1238 
1239 		/* Invalidate the RX mbuf and unload its map */
1240 
1241 		bus_dmamap_sync(sc->sc_dmat, rxs->rxs_dmamap,
1242 		    0, rxs->rxs_dmamap->dm_mapsize, BUS_DMASYNC_POSTREAD);
1243 		bus_dmamap_unload(sc->sc_dmat, rxs->rxs_dmamap);
1244 
1245 		/*
1246 		 * If the 'start of frame' bit is set, this indicates
1247 		 * either the first fragment in a multi-fragment receive,
1248 		 * or an intermediate fragment. Either way, we want to
1249 		 * accumulate the buffers.
1250 		 */
1251 		if (rxstat & VGE_RXPKT_SOF) {
1252 			m->m_len = VGE_RX_BUFSIZE;
1253 			if (sc->sc_rx_mhead == NULL)
1254 				sc->sc_rx_mhead = sc->sc_rx_mtail = m;
1255 			else {
1256 				m->m_flags &= ~M_PKTHDR;
1257 				sc->sc_rx_mtail->m_next = m;
1258 				sc->sc_rx_mtail = m;
1259 			}
1260 			vge_newbuf(sc, idx, NULL);
1261 			continue;
1262 		}
1263 
1264 		/*
1265 		 * Bad/error frames will have the RXOK bit cleared.
1266 		 * However, there's one error case we want to allow:
1267 		 * if a VLAN tagged frame arrives and the chip can't
1268 		 * match it against the CAM filter, it considers this
1269 		 * a 'VLAN CAM filter miss' and clears the 'RXOK' bit.
1270 		 * We don't want to drop the frame though: our VLAN
1271 		 * filtering is done in software.
1272 		 */
1273 		if ((rxstat & VGE_RDSTS_RXOK) == 0 &&
1274 		    (rxstat & VGE_RDSTS_VIDM) == 0 &&
1275 		    (rxstat & VGE_RDSTS_CSUMERR) == 0) {
1276 			ifp->if_ierrors++;
1277 			/*
1278 			 * If this is part of a multi-fragment packet,
1279 			 * discard all the pieces.
1280 			 */
1281 			if (sc->sc_rx_mhead != NULL) {
1282 				m_freem(sc->sc_rx_mhead);
1283 				sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1284 			}
1285 			vge_newbuf(sc, idx, m);
1286 			continue;
1287 		}
1288 
1289 		/*
1290 		 * If allocating a replacement mbuf fails,
1291 		 * reload the current one.
1292 		 */
1293 
1294 		if (vge_newbuf(sc, idx, NULL)) {
1295 			ifp->if_ierrors++;
1296 			if (sc->sc_rx_mhead != NULL) {
1297 				m_freem(sc->sc_rx_mhead);
1298 				sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1299 			}
1300 			vge_newbuf(sc, idx, m);
1301 			continue;
1302 		}
1303 
1304 		if (sc->sc_rx_mhead != NULL) {
1305 			m->m_len = total_len % VGE_RX_BUFSIZE;
1306 			/*
1307 			 * Special case: if there's 4 bytes or less
1308 			 * in this buffer, the mbuf can be discarded:
1309 			 * the last 4 bytes is the CRC, which we don't
1310 			 * care about anyway.
1311 			 */
1312 			if (m->m_len <= ETHER_CRC_LEN) {
1313 				sc->sc_rx_mtail->m_len -=
1314 				    (ETHER_CRC_LEN - m->m_len);
1315 				m_freem(m);
1316 			} else {
1317 				m->m_len -= ETHER_CRC_LEN;
1318 				m->m_flags &= ~M_PKTHDR;
1319 				sc->sc_rx_mtail->m_next = m;
1320 			}
1321 			m = sc->sc_rx_mhead;
1322 			sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1323 			m->m_pkthdr.len = total_len - ETHER_CRC_LEN;
1324 		} else
1325 			m->m_pkthdr.len = m->m_len = total_len - ETHER_CRC_LEN;
1326 
1327 #ifndef __NO_STRICT_ALIGNMENT
1328 		vge_fixup_rx(m);
1329 #endif
1330 		ifp->if_ipackets++;
1331 		m->m_pkthdr.rcvif = ifp;
1332 
1333 		/* Do RX checksumming if enabled */
1334 		if (ifp->if_csum_flags_rx & M_CSUM_IPv4) {
1335 
1336 			/* Check IP header checksum */
1337 			if (rxctl & VGE_RDCTL_IPPKT)
1338 				m->m_pkthdr.csum_flags |= M_CSUM_IPv4;
1339 			if ((rxctl & VGE_RDCTL_IPCSUMOK) == 0)
1340 				m->m_pkthdr.csum_flags |= M_CSUM_IPv4_BAD;
1341 		}
1342 
1343 		if (ifp->if_csum_flags_rx & M_CSUM_TCPv4) {
1344 			/* Check UDP checksum */
1345 			if (rxctl & VGE_RDCTL_TCPPKT)
1346 				m->m_pkthdr.csum_flags |= M_CSUM_TCPv4;
1347 
1348 			if ((rxctl & VGE_RDCTL_PROTOCSUMOK) == 0)
1349 				m->m_pkthdr.csum_flags |= M_CSUM_TCP_UDP_BAD;
1350 		}
1351 
1352 		if (ifp->if_csum_flags_rx & M_CSUM_UDPv4) {
1353 			/* Check UDP checksum */
1354 			if (rxctl & VGE_RDCTL_UDPPKT)
1355 				m->m_pkthdr.csum_flags |= M_CSUM_UDPv4;
1356 
1357 			if ((rxctl & VGE_RDCTL_PROTOCSUMOK) == 0)
1358 				m->m_pkthdr.csum_flags |= M_CSUM_TCP_UDP_BAD;
1359 		}
1360 
1361 		if (rxstat & VGE_RDSTS_VTAG) {
1362 			/*
1363 			 * We use bswap16() here because:
1364 			 * On LE machines, tag is stored in BE as stream data.
1365 			 * On BE machines, tag is stored in BE as stream data
1366 			 *  but it was already swapped by le32toh() above.
1367 			 */
1368 			VLAN_INPUT_TAG(ifp, m,
1369 			    bswap16(rxctl & VGE_RDCTL_VLANID), continue);
1370 		}
1371 
1372 #if NBPFILTER > 0
1373 		/*
1374 		 * Handle BPF listeners.
1375 		 */
1376 		if (ifp->if_bpf)
1377 			bpf_mtap(ifp->if_bpf, m);
1378 #endif
1379 
1380 		(*ifp->if_input)(ifp, m);
1381 
1382 		lim++;
1383 		if (lim == VGE_NRXDESC)
1384 			break;
1385 	}
1386 
1387 	sc->sc_rx_prodidx = idx;
1388 	CSR_WRITE_2(sc, VGE_RXDESC_RESIDUECNT, lim);
1389 }
1390 
1391 static void
1392 vge_txeof(struct vge_softc *sc)
1393 {
1394 	struct ifnet *ifp;
1395 	struct vge_txsoft *txs;
1396 	uint32_t txstat;
1397 	int idx;
1398 
1399 	ifp = &sc->sc_ethercom.ec_if;
1400 
1401 	for (idx = sc->sc_tx_considx;
1402 	    sc->sc_tx_free < VGE_NTXDESC;
1403 	    idx = VGE_NEXT_TXDESC(idx), sc->sc_tx_free++) {
1404 		VGE_TXDESCSYNC(sc, idx,
1405 		    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1406 		txstat = le32toh(sc->sc_txdescs[idx].td_sts);
1407 		VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1408 		if (txstat & VGE_TDSTS_OWN) {
1409 			break;
1410 		}
1411 
1412 		txs = &sc->sc_txsoft[idx];
1413 		m_freem(txs->txs_mbuf);
1414 		txs->txs_mbuf = NULL;
1415 		bus_dmamap_sync(sc->sc_dmat, txs->txs_dmamap, 0,
1416 		    txs->txs_dmamap->dm_mapsize, BUS_DMASYNC_POSTWRITE);
1417 		bus_dmamap_unload(sc->sc_dmat, txs->txs_dmamap);
1418 		if (txstat & (VGE_TDSTS_EXCESSCOLL|VGE_TDSTS_COLL))
1419 			ifp->if_collisions++;
1420 		if (txstat & VGE_TDSTS_TXERR)
1421 			ifp->if_oerrors++;
1422 		else
1423 			ifp->if_opackets++;
1424 	}
1425 
1426 	sc->sc_tx_considx = idx;
1427 
1428 	if (sc->sc_tx_free > 0) {
1429 		ifp->if_flags &= ~IFF_OACTIVE;
1430 	}
1431 
1432 	/*
1433 	 * If not all descriptors have been released reaped yet,
1434 	 * reload the timer so that we will eventually get another
1435 	 * interrupt that will cause us to re-enter this routine.
1436 	 * This is done in case the transmitter has gone idle.
1437 	 */
1438 	if (sc->sc_tx_free < VGE_NTXDESC)
1439 		CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_TIMER0_ENABLE);
1440 	else
1441 		ifp->if_timer = 0;
1442 }
1443 
1444 static void
1445 vge_tick(void *arg)
1446 {
1447 	struct vge_softc *sc;
1448 	struct ifnet *ifp;
1449 	struct mii_data *mii;
1450 	int s;
1451 
1452 	sc = arg;
1453 	ifp = &sc->sc_ethercom.ec_if;
1454 	mii = &sc->sc_mii;
1455 
1456 	s = splnet();
1457 
1458 	callout_schedule(&sc->sc_timeout, hz);
1459 
1460 	mii_tick(mii);
1461 	if (sc->sc_link) {
1462 		if ((mii->mii_media_status & IFM_ACTIVE) == 0)
1463 			sc->sc_link = 0;
1464 	} else {
1465 		if (mii->mii_media_status & IFM_ACTIVE &&
1466 		    IFM_SUBTYPE(mii->mii_media_active) != IFM_NONE) {
1467 			sc->sc_link = 1;
1468 			if (!IFQ_IS_EMPTY(&ifp->if_snd))
1469 				vge_start(ifp);
1470 		}
1471 	}
1472 
1473 	splx(s);
1474 }
1475 
1476 static int
1477 vge_intr(void *arg)
1478 {
1479 	struct vge_softc *sc;
1480 	struct ifnet *ifp;
1481 	uint32_t status;
1482 	int claim;
1483 
1484 	sc = arg;
1485 	claim = 0;
1486 	if (sc->sc_suspended) {
1487 		return claim;
1488 	}
1489 
1490 	ifp = &sc->sc_ethercom.ec_if;
1491 
1492 	if ((ifp->if_flags & IFF_UP) == 0) {
1493 		return claim;
1494 	}
1495 
1496 	/* Disable interrupts */
1497 	CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
1498 
1499 	for (;;) {
1500 
1501 		status = CSR_READ_4(sc, VGE_ISR);
1502 		/* If the card has gone away the read returns 0xffffffff. */
1503 		if (status == 0xFFFFFFFF)
1504 			break;
1505 
1506 		if (status) {
1507 			claim = 1;
1508 			CSR_WRITE_4(sc, VGE_ISR, status);
1509 		}
1510 
1511 		if ((status & VGE_INTRS) == 0)
1512 			break;
1513 
1514 		if (status & (VGE_ISR_RXOK|VGE_ISR_RXOK_HIPRIO))
1515 			vge_rxeof(sc);
1516 
1517 		if (status & (VGE_ISR_RXOFLOW|VGE_ISR_RXNODESC)) {
1518 			vge_rxeof(sc);
1519 			CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_RUN);
1520 			CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_WAK);
1521 		}
1522 
1523 		if (status & (VGE_ISR_TXOK0|VGE_ISR_TIMER0))
1524 			vge_txeof(sc);
1525 
1526 		if (status & (VGE_ISR_TXDMA_STALL|VGE_ISR_RXDMA_STALL))
1527 			vge_init(ifp);
1528 
1529 		if (status & VGE_ISR_LINKSTS)
1530 			vge_tick(sc);
1531 	}
1532 
1533 	/* Re-enable interrupts */
1534 	CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_GMSK);
1535 
1536 	if (claim && !IFQ_IS_EMPTY(&ifp->if_snd))
1537 		vge_start(ifp);
1538 
1539 	return claim;
1540 }
1541 
1542 static int
1543 vge_encap(struct vge_softc *sc, struct mbuf *m_head, int idx)
1544 {
1545 	struct vge_txsoft *txs;
1546 	struct vge_txdesc *txd;
1547 	struct vge_txfrag *f;
1548 	struct mbuf *m_new;
1549 	bus_dmamap_t map;
1550 	int m_csumflags, seg, error, flags;
1551 	struct m_tag *mtag;
1552 	size_t sz;
1553 	uint32_t td_sts, td_ctl;
1554 
1555 	KASSERT(sc->sc_tx_free > 0);
1556 
1557 	txd = &sc->sc_txdescs[idx];
1558 
1559 #ifdef DIAGNOSTIC
1560 	/* If this descriptor is still owned by the chip, bail. */
1561 	VGE_TXDESCSYNC(sc, idx,
1562 	    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1563 	td_sts = le32toh(txd->td_sts);
1564 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1565 	if (td_sts & VGE_TDSTS_OWN) {
1566 		return ENOBUFS;
1567 	}
1568 #endif
1569 
1570 	/*
1571 	 * Preserve m_pkthdr.csum_flags here since m_head might be
1572 	 * updated by m_defrag()
1573 	 */
1574 	m_csumflags = m_head->m_pkthdr.csum_flags;
1575 
1576 	txs = &sc->sc_txsoft[idx];
1577 	map = txs->txs_dmamap;
1578 	error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m_head, BUS_DMA_NOWAIT);
1579 
1580 	/* If too many segments to map, coalesce */
1581 	if (error == EFBIG ||
1582 	    (m_head->m_pkthdr.len < ETHER_PAD_LEN &&
1583 	     map->dm_nsegs == VGE_TX_FRAGS)) {
1584 		m_new = m_defrag(m_head, M_DONTWAIT);
1585 		if (m_new == NULL)
1586 			return EFBIG;
1587 
1588 		error = bus_dmamap_load_mbuf(sc->sc_dmat, map,
1589 		    m_new, BUS_DMA_NOWAIT);
1590 		if (error) {
1591 			m_freem(m_new);
1592 			return error;
1593 		}
1594 
1595 		m_head = m_new;
1596 	} else if (error)
1597 		return error;
1598 
1599 	txs->txs_mbuf = m_head;
1600 
1601 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1602 	    BUS_DMASYNC_PREWRITE);
1603 
1604 	for (seg = 0, f = &txd->td_frag[0]; seg < map->dm_nsegs; seg++, f++) {
1605 		f->tf_buflen = htole16(VGE_BUFLEN(map->dm_segs[seg].ds_len));
1606 		vge_set_txaddr(f, map->dm_segs[seg].ds_addr);
1607 	}
1608 
1609 	/* Argh. This chip does not autopad short frames */
1610 	sz = m_head->m_pkthdr.len;
1611 	if (sz < ETHER_PAD_LEN) {
1612 		f->tf_buflen = htole16(VGE_BUFLEN(ETHER_PAD_LEN - sz));
1613 		vge_set_txaddr(f, VGE_CDPADADDR(sc));
1614 		sz = ETHER_PAD_LEN;
1615 		seg++;
1616 	}
1617 	VGE_TXFRAGSYNC(sc, idx, seg, BUS_DMASYNC_PREWRITE);
1618 
1619 	/*
1620 	 * When telling the chip how many segments there are, we
1621 	 * must use nsegs + 1 instead of just nsegs. Darned if I
1622 	 * know why.
1623 	 */
1624 	seg++;
1625 
1626 	flags = 0;
1627 	if (m_csumflags & M_CSUM_IPv4)
1628 		flags |= VGE_TDCTL_IPCSUM;
1629 	if (m_csumflags & M_CSUM_TCPv4)
1630 		flags |= VGE_TDCTL_TCPCSUM;
1631 	if (m_csumflags & M_CSUM_UDPv4)
1632 		flags |= VGE_TDCTL_UDPCSUM;
1633 	td_sts = sz << 16;
1634 	td_ctl = flags | (seg << 28) | VGE_TD_LS_NORM;
1635 
1636 	if (sz > ETHERMTU + ETHER_HDR_LEN)
1637 		td_ctl |= VGE_TDCTL_JUMBO;
1638 
1639 	/*
1640 	 * Set up hardware VLAN tagging.
1641 	 */
1642 	mtag = VLAN_OUTPUT_TAG(&sc->sc_ethercom, m_head);
1643 	if (mtag != NULL) {
1644 		/*
1645 		 * No need htons() here since vge(4) chip assumes
1646 		 * that tags are written in little endian and
1647 		 * we already use htole32() here.
1648 		 */
1649 		td_ctl |= VLAN_TAG_VALUE(mtag) | VGE_TDCTL_VTAG;
1650 	}
1651 	txd->td_ctl = htole32(td_ctl);
1652 	txd->td_sts = htole32(td_sts);
1653 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1654 
1655 	txd->td_sts = htole32(VGE_TDSTS_OWN | td_sts);
1656 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1657 
1658 	sc->sc_tx_free--;
1659 
1660 	return 0;
1661 }
1662 
1663 /*
1664  * Main transmit routine.
1665  */
1666 
1667 static void
1668 vge_start(struct ifnet *ifp)
1669 {
1670 	struct vge_softc *sc;
1671 	struct vge_txsoft *txs;
1672 	struct mbuf *m_head;
1673 	int idx, pidx, ofree, error;
1674 
1675 	sc = ifp->if_softc;
1676 
1677 	if (!sc->sc_link ||
1678 	    (ifp->if_flags & (IFF_RUNNING|IFF_OACTIVE)) != IFF_RUNNING) {
1679 		return;
1680 	}
1681 
1682 	m_head = NULL;
1683 	idx = sc->sc_tx_prodidx;
1684 	pidx = VGE_PREV_TXDESC(idx);
1685 	ofree = sc->sc_tx_free;
1686 
1687 	/*
1688 	 * Loop through the send queue, setting up transmit descriptors
1689 	 * until we drain the queue, or use up all available transmit
1690 	 * descriptors.
1691 	 */
1692 	for (;;) {
1693 		/* Grab a packet off the queue. */
1694 		IFQ_POLL(&ifp->if_snd, m_head);
1695 		if (m_head == NULL)
1696 			break;
1697 
1698 		if (sc->sc_tx_free == 0) {
1699 			/*
1700 			 * All slots used, stop for now.
1701 			 */
1702 			ifp->if_flags |= IFF_OACTIVE;
1703 			break;
1704 		}
1705 
1706 		txs = &sc->sc_txsoft[idx];
1707 		KASSERT(txs->txs_mbuf == NULL);
1708 
1709 		if ((error = vge_encap(sc, m_head, idx))) {
1710 			if (error == EFBIG) {
1711 				printf("%s: Tx packet consumes too many "
1712 				    "DMA segments, dropping...\n",
1713 				    device_xname(sc->sc_dev));
1714 				IFQ_DEQUEUE(&ifp->if_snd, m_head);
1715 				m_freem(m_head);
1716 				continue;
1717 			}
1718 
1719 			/*
1720 			 * Short on resources, just stop for now.
1721 			 */
1722 			if (error == ENOBUFS)
1723 				ifp->if_flags |= IFF_OACTIVE;
1724 			break;
1725 		}
1726 
1727 		IFQ_DEQUEUE(&ifp->if_snd, m_head);
1728 
1729 		/*
1730 		 * WE ARE NOW COMMITTED TO TRANSMITTING THE PACKET.
1731 		 */
1732 
1733 		sc->sc_txdescs[pidx].td_frag[0].tf_buflen |=
1734 		    htole16(VGE_TXDESC_Q);
1735 		VGE_TXFRAGSYNC(sc, pidx, 1,
1736 		    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1737 
1738 		if (txs->txs_mbuf != m_head) {
1739 			m_freem(m_head);
1740 			m_head = txs->txs_mbuf;
1741 		}
1742 
1743 		pidx = idx;
1744 		idx = VGE_NEXT_TXDESC(idx);
1745 
1746 		/*
1747 		 * If there's a BPF listener, bounce a copy of this frame
1748 		 * to him.
1749 		 */
1750 #if NBPFILTER > 0
1751 		if (ifp->if_bpf)
1752 			bpf_mtap(ifp->if_bpf, m_head);
1753 #endif
1754 	}
1755 
1756 	if (sc->sc_tx_free < ofree) {
1757 		/* TX packet queued */
1758 
1759 		sc->sc_tx_prodidx = idx;
1760 
1761 		/* Issue a transmit command. */
1762 		CSR_WRITE_2(sc, VGE_TXQCSRS, VGE_TXQCSR_WAK0);
1763 
1764 		/*
1765 		 * Use the countdown timer for interrupt moderation.
1766 		 * 'TX done' interrupts are disabled. Instead, we reset the
1767 		 * countdown timer, which will begin counting until it hits
1768 		 * the value in the SSTIMER register, and then trigger an
1769 		 * interrupt. Each time we set the TIMER0_ENABLE bit, the
1770 		 * the timer count is reloaded. Only when the transmitter
1771 		 * is idle will the timer hit 0 and an interrupt fire.
1772 		 */
1773 		CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_TIMER0_ENABLE);
1774 
1775 		/*
1776 		 * Set a timeout in case the chip goes out to lunch.
1777 		 */
1778 		ifp->if_timer = 5;
1779 	}
1780 }
1781 
1782 static int
1783 vge_init(struct ifnet *ifp)
1784 {
1785 	struct vge_softc *sc;
1786 	int i, rc = 0;
1787 
1788 	sc = ifp->if_softc;
1789 
1790 	/*
1791 	 * Cancel pending I/O and free all RX/TX buffers.
1792 	 */
1793 	vge_stop(ifp, 0);
1794 	vge_reset(sc);
1795 
1796 	/* Initialize the RX descriptors and mbufs. */
1797 	memset(sc->sc_rxdescs, 0, sizeof(sc->sc_rxdescs));
1798 	sc->sc_rx_consumed = 0;
1799 	for (i = 0; i < VGE_NRXDESC; i++) {
1800 		if (vge_newbuf(sc, i, NULL) == ENOBUFS) {
1801 			printf("%s: unable to allocate or map rx buffer\n",
1802 			    device_xname(sc->sc_dev));
1803 			return 1; /* XXX */
1804 		}
1805 	}
1806 	sc->sc_rx_prodidx = 0;
1807 	sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1808 
1809 	/* Initialize the  TX descriptors and mbufs. */
1810 	memset(sc->sc_txdescs, 0, sizeof(sc->sc_txdescs));
1811 	bus_dmamap_sync(sc->sc_dmat, sc->sc_cddmamap,
1812 	    VGE_CDTXOFF(0), sizeof(sc->sc_txdescs),
1813 	    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1814 	for (i = 0; i < VGE_NTXDESC; i++)
1815 		sc->sc_txsoft[i].txs_mbuf = NULL;
1816 
1817 	sc->sc_tx_prodidx = 0;
1818 	sc->sc_tx_considx = 0;
1819 	sc->sc_tx_free = VGE_NTXDESC;
1820 
1821 	/* Set our station address */
1822 	for (i = 0; i < ETHER_ADDR_LEN; i++)
1823 		CSR_WRITE_1(sc, VGE_PAR0 + i, sc->sc_eaddr[i]);
1824 
1825 	/*
1826 	 * Set receive FIFO threshold. Also allow transmission and
1827 	 * reception of VLAN tagged frames.
1828 	 */
1829 	CSR_CLRBIT_1(sc, VGE_RXCFG, VGE_RXCFG_FIFO_THR|VGE_RXCFG_VTAGOPT);
1830 	CSR_SETBIT_1(sc, VGE_RXCFG, VGE_RXFIFOTHR_128BYTES|VGE_VTAG_OPT2);
1831 
1832 	/* Set DMA burst length */
1833 	CSR_CLRBIT_1(sc, VGE_DMACFG0, VGE_DMACFG0_BURSTLEN);
1834 	CSR_SETBIT_1(sc, VGE_DMACFG0, VGE_DMABURST_128);
1835 
1836 	CSR_SETBIT_1(sc, VGE_TXCFG, VGE_TXCFG_ARB_PRIO|VGE_TXCFG_NONBLK);
1837 
1838 	/* Set collision backoff algorithm */
1839 	CSR_CLRBIT_1(sc, VGE_CHIPCFG1, VGE_CHIPCFG1_CRANDOM|
1840 	    VGE_CHIPCFG1_CAP|VGE_CHIPCFG1_MBA|VGE_CHIPCFG1_BAKOPT);
1841 	CSR_SETBIT_1(sc, VGE_CHIPCFG1, VGE_CHIPCFG1_OFSET);
1842 
1843 	/* Disable LPSEL field in priority resolution */
1844 	CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_LPSEL_DIS);
1845 
1846 	/*
1847 	 * Load the addresses of the DMA queues into the chip.
1848 	 * Note that we only use one transmit queue.
1849 	 */
1850 
1851 	CSR_WRITE_4(sc, VGE_TXDESC_ADDR_LO0, VGE_ADDR_LO(VGE_CDTXADDR(sc, 0)));
1852 	CSR_WRITE_2(sc, VGE_TXDESCNUM, VGE_NTXDESC - 1);
1853 
1854 	CSR_WRITE_4(sc, VGE_RXDESC_ADDR_LO, VGE_ADDR_LO(VGE_CDRXADDR(sc, 0)));
1855 	CSR_WRITE_2(sc, VGE_RXDESCNUM, VGE_NRXDESC - 1);
1856 	CSR_WRITE_2(sc, VGE_RXDESC_RESIDUECNT, VGE_NRXDESC);
1857 
1858 	/* Enable and wake up the RX descriptor queue */
1859 	CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_RUN);
1860 	CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_WAK);
1861 
1862 	/* Enable the TX descriptor queue */
1863 	CSR_WRITE_2(sc, VGE_TXQCSRS, VGE_TXQCSR_RUN0);
1864 
1865 	/* Set up the receive filter -- allow large frames for VLANs. */
1866 	CSR_WRITE_1(sc, VGE_RXCTL, VGE_RXCTL_RX_UCAST|VGE_RXCTL_RX_GIANT);
1867 
1868 	/* If we want promiscuous mode, set the allframes bit. */
1869 	if (ifp->if_flags & IFF_PROMISC) {
1870 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_PROMISC);
1871 	}
1872 
1873 	/* Set capture broadcast bit to capture broadcast frames. */
1874 	if (ifp->if_flags & IFF_BROADCAST) {
1875 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_BCAST);
1876 	}
1877 
1878 	/* Set multicast bit to capture multicast frames. */
1879 	if (ifp->if_flags & IFF_MULTICAST) {
1880 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_MCAST);
1881 	}
1882 
1883 	/* Init the cam filter. */
1884 	vge_cam_clear(sc);
1885 
1886 	/* Init the multicast filter. */
1887 	vge_setmulti(sc);
1888 
1889 	/* Enable flow control */
1890 
1891 	CSR_WRITE_1(sc, VGE_CRS2, 0x8B);
1892 
1893 	/* Enable jumbo frame reception (if desired) */
1894 
1895 	/* Start the MAC. */
1896 	CSR_WRITE_1(sc, VGE_CRC0, VGE_CR0_STOP);
1897 	CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_NOPOLL);
1898 	CSR_WRITE_1(sc, VGE_CRS0,
1899 	    VGE_CR0_TX_ENABLE|VGE_CR0_RX_ENABLE|VGE_CR0_START);
1900 
1901 	/*
1902 	 * Configure one-shot timer for microsecond
1903 	 * resulution and load it for 500 usecs.
1904 	 */
1905 	CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_TIMER0_RES);
1906 	CSR_WRITE_2(sc, VGE_SSTIMER, 400);
1907 
1908 	/*
1909 	 * Configure interrupt moderation for receive. Enable
1910 	 * the holdoff counter and load it, and set the RX
1911 	 * suppression count to the number of descriptors we
1912 	 * want to allow before triggering an interrupt.
1913 	 * The holdoff timer is in units of 20 usecs.
1914 	 */
1915 
1916 #ifdef notyet
1917 	CSR_WRITE_1(sc, VGE_INTCTL1, VGE_INTCTL_TXINTSUP_DISABLE);
1918 	/* Select the interrupt holdoff timer page. */
1919 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1920 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_INTHLDOFF);
1921 	CSR_WRITE_1(sc, VGE_INTHOLDOFF, 10); /* ~200 usecs */
1922 
1923 	/* Enable use of the holdoff timer. */
1924 	CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_HOLDOFF);
1925 	CSR_WRITE_1(sc, VGE_INTCTL1, VGE_INTCTL_SC_RELOAD);
1926 
1927 	/* Select the RX suppression threshold page. */
1928 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1929 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_RXSUPPTHR);
1930 	CSR_WRITE_1(sc, VGE_RXSUPPTHR, 64); /* interrupt after 64 packets */
1931 
1932 	/* Restore the page select bits. */
1933 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1934 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_MAR);
1935 #endif
1936 
1937 #ifdef DEVICE_POLLING
1938 	/*
1939 	 * Disable interrupts if we are polling.
1940 	 */
1941 	if (ifp->if_flags & IFF_POLLING) {
1942 		CSR_WRITE_4(sc, VGE_IMR, 0);
1943 		CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
1944 	} else	/* otherwise ... */
1945 #endif /* DEVICE_POLLING */
1946 	{
1947 	/*
1948 	 * Enable interrupts.
1949 	 */
1950 		CSR_WRITE_4(sc, VGE_IMR, VGE_INTRS);
1951 		CSR_WRITE_4(sc, VGE_ISR, 0);
1952 		CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_GMSK);
1953 	}
1954 
1955 	if ((rc = ether_mediachange(ifp)) != 0)
1956 		goto out;
1957 
1958 	ifp->if_flags |= IFF_RUNNING;
1959 	ifp->if_flags &= ~IFF_OACTIVE;
1960 
1961 	sc->sc_if_flags = 0;
1962 	sc->sc_link = 0;
1963 
1964 	callout_schedule(&sc->sc_timeout, hz);
1965 
1966 out:
1967 	return rc;
1968 }
1969 
1970 static void
1971 vge_miibus_statchg(device_t self)
1972 {
1973 	struct vge_softc *sc;
1974 	struct mii_data *mii;
1975 	struct ifmedia_entry *ife;
1976 
1977 	sc = device_private(self);
1978 	mii = &sc->sc_mii;
1979 	ife = mii->mii_media.ifm_cur;
1980 	/*
1981 	 * If the user manually selects a media mode, we need to turn
1982 	 * on the forced MAC mode bit in the DIAGCTL register. If the
1983 	 * user happens to choose a full duplex mode, we also need to
1984 	 * set the 'force full duplex' bit. This applies only to
1985 	 * 10Mbps and 100Mbps speeds. In autoselect mode, forced MAC
1986 	 * mode is disabled, and in 1000baseT mode, full duplex is
1987 	 * always implied, so we turn on the forced mode bit but leave
1988 	 * the FDX bit cleared.
1989 	 */
1990 
1991 	switch (IFM_SUBTYPE(ife->ifm_media)) {
1992 	case IFM_AUTO:
1993 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
1994 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1995 		break;
1996 	case IFM_1000_T:
1997 		CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
1998 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1999 		break;
2000 	case IFM_100_TX:
2001 	case IFM_10_T:
2002 		CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
2003 		if ((ife->ifm_media & IFM_GMASK) == IFM_FDX) {
2004 			CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
2005 		} else {
2006 			CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
2007 		}
2008 		break;
2009 	default:
2010 		printf("%s: unknown media type: %x\n",
2011 		    device_xname(sc->sc_dev),
2012 		    IFM_SUBTYPE(ife->ifm_media));
2013 		break;
2014 	}
2015 }
2016 
2017 static int
2018 vge_ifflags_cb(struct ethercom *ec)
2019 {
2020 	struct ifnet *ifp = &ec->ec_if;
2021 	struct vge_softc *sc = ifp->if_softc;
2022 	int change = ifp->if_flags ^ sc->sc_if_flags;
2023 
2024 	if ((change & ~(IFF_CANTCHANGE|IFF_DEBUG)) != 0)
2025 		return ENETRESET;
2026 	else if ((change & IFF_PROMISC) == 0)
2027 		return 0;
2028 
2029 	if ((ifp->if_flags & IFF_PROMISC) == 0)
2030 		CSR_CLRBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_PROMISC);
2031 	else
2032 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_PROMISC);
2033 	vge_setmulti(sc);
2034 	return 0;
2035 }
2036 
2037 static int
2038 vge_ioctl(struct ifnet *ifp, u_long command, void *data)
2039 {
2040 	struct vge_softc *sc;
2041 	struct ifreq *ifr;
2042 	int s, error;
2043 
2044 	sc = ifp->if_softc;
2045 	ifr = (struct ifreq *)data;
2046 	error = 0;
2047 
2048 	s = splnet();
2049 
2050 	if ((error = ether_ioctl(ifp, command, data)) == ENETRESET) {
2051 		error = 0;
2052 		if (command != SIOCADDMULTI && command != SIOCDELMULTI)
2053 			;
2054 		else if (ifp->if_flags & IFF_RUNNING) {
2055 			/*
2056 			 * Multicast list has changed; set the hardware filter
2057 			 * accordingly.
2058 			 */
2059 			vge_setmulti(sc);
2060 		}
2061 	}
2062 	sc->sc_if_flags = ifp->if_flags;
2063 
2064 	splx(s);
2065 	return error;
2066 }
2067 
2068 static void
2069 vge_watchdog(struct ifnet *ifp)
2070 {
2071 	struct vge_softc *sc;
2072 	int s;
2073 
2074 	sc = ifp->if_softc;
2075 	s = splnet();
2076 	printf("%s: watchdog timeout\n", device_xname(sc->sc_dev));
2077 	ifp->if_oerrors++;
2078 
2079 	vge_txeof(sc);
2080 	vge_rxeof(sc);
2081 
2082 	vge_init(ifp);
2083 
2084 	splx(s);
2085 }
2086 
2087 /*
2088  * Stop the adapter and free any mbufs allocated to the
2089  * RX and TX lists.
2090  */
2091 static void
2092 vge_stop(struct ifnet *ifp, int disable)
2093 {
2094 	struct vge_softc *sc = ifp->if_softc;
2095 	struct vge_txsoft *txs;
2096 	struct vge_rxsoft *rxs;
2097 	int i, s;
2098 
2099 	s = splnet();
2100 	ifp->if_timer = 0;
2101 
2102 	ifp->if_flags &= ~(IFF_RUNNING | IFF_OACTIVE);
2103 #ifdef DEVICE_POLLING
2104 	ether_poll_deregister(ifp);
2105 #endif /* DEVICE_POLLING */
2106 
2107 	CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
2108 	CSR_WRITE_1(sc, VGE_CRS0, VGE_CR0_STOP);
2109 	CSR_WRITE_4(sc, VGE_ISR, 0xFFFFFFFF);
2110 	CSR_WRITE_2(sc, VGE_TXQCSRC, 0xFFFF);
2111 	CSR_WRITE_1(sc, VGE_RXQCSRC, 0xFF);
2112 	CSR_WRITE_4(sc, VGE_RXDESC_ADDR_LO, 0);
2113 
2114 	if (sc->sc_rx_mhead != NULL) {
2115 		m_freem(sc->sc_rx_mhead);
2116 		sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
2117 	}
2118 
2119 	/* Free the TX list buffers. */
2120 
2121 	for (i = 0; i < VGE_NTXDESC; i++) {
2122 		txs = &sc->sc_txsoft[i];
2123 		if (txs->txs_mbuf != NULL) {
2124 			bus_dmamap_unload(sc->sc_dmat, txs->txs_dmamap);
2125 			m_freem(txs->txs_mbuf);
2126 			txs->txs_mbuf = NULL;
2127 		}
2128 	}
2129 
2130 	/* Free the RX list buffers. */
2131 
2132 	for (i = 0; i < VGE_NRXDESC; i++) {
2133 		rxs = &sc->sc_rxsoft[i];
2134 		if (rxs->rxs_mbuf != NULL) {
2135 			bus_dmamap_unload(sc->sc_dmat, rxs->rxs_dmamap);
2136 			m_freem(rxs->rxs_mbuf);
2137 			rxs->rxs_mbuf = NULL;
2138 		}
2139 	}
2140 
2141 	splx(s);
2142 }
2143 
2144 #if VGE_POWER_MANAGEMENT
2145 /*
2146  * Device suspend routine.  Stop the interface and save some PCI
2147  * settings in case the BIOS doesn't restore them properly on
2148  * resume.
2149  */
2150 static int
2151 vge_suspend(device_t dev)
2152 {
2153 	struct vge_softc *sc;
2154 	int i;
2155 
2156 	sc = device_get_softc(dev);
2157 
2158 	vge_stop(sc);
2159 
2160         for (i = 0; i < 5; i++)
2161 		sc->sc_saved_maps[i] =
2162 		    pci_read_config(dev, PCIR_MAPS + i * 4, 4);
2163 	sc->sc_saved_biosaddr = pci_read_config(dev, PCIR_BIOS, 4);
2164 	sc->sc_saved_intline = pci_read_config(dev, PCIR_INTLINE, 1);
2165 	sc->sc_saved_cachelnsz = pci_read_config(dev, PCIR_CACHELNSZ, 1);
2166 	sc->sc_saved_lattimer = pci_read_config(dev, PCIR_LATTIMER, 1);
2167 
2168 	sc->suspended = 1;
2169 
2170 	return 0;
2171 }
2172 
2173 /*
2174  * Device resume routine.  Restore some PCI settings in case the BIOS
2175  * doesn't, re-enable busmastering, and restart the interface if
2176  * appropriate.
2177  */
2178 static int
2179 vge_resume(device_t dev)
2180 {
2181 	struct vge_softc *sc;
2182 	struct ifnet *ifp;
2183 	int i;
2184 
2185 	sc = device_private(dev);
2186 	ifp = &sc->sc_ethercom.ec_if;
2187 
2188         /* better way to do this? */
2189 	for (i = 0; i < 5; i++)
2190 		pci_write_config(dev, PCIR_MAPS + i * 4,
2191 		    sc->sc_saved_maps[i], 4);
2192 	pci_write_config(dev, PCIR_BIOS, sc->sc_saved_biosaddr, 4);
2193 	pci_write_config(dev, PCIR_INTLINE, sc->sc_saved_intline, 1);
2194 	pci_write_config(dev, PCIR_CACHELNSZ, sc->sc_saved_cachelnsz, 1);
2195 	pci_write_config(dev, PCIR_LATTIMER, sc->sc_saved_lattimer, 1);
2196 
2197 	/* reenable busmastering */
2198 	pci_enable_busmaster(dev);
2199 	pci_enable_io(dev, SYS_RES_MEMORY);
2200 
2201 	/* reinitialize interface if necessary */
2202 	if (ifp->if_flags & IFF_UP)
2203 		vge_init(sc);
2204 
2205 	sc->suspended = 0;
2206 
2207 	return 0;
2208 }
2209 #endif
2210 
2211 /*
2212  * Stop all chip I/O so that the kernel's probe routines don't
2213  * get confused by errant DMAs when rebooting.
2214  */
2215 static bool
2216 vge_shutdown(device_t self, int howto)
2217 {
2218 	struct vge_softc *sc;
2219 
2220 	sc = device_private(self);
2221 	vge_stop(&sc->sc_ethercom.ec_if, 1);
2222 
2223 	return true;
2224 }
2225