xref: /netbsd-src/sys/dev/pci/if_vge.c (revision e89934bbf778a6d6d6894877c4da59d0c7835b0f)
1 /* $NetBSD: if_vge.c,v 1.60 2016/12/15 09:28:05 ozaki-r 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 1.60 2016/12/15 09:28:05 ozaki-r 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 
88 #include <sys/param.h>
89 #include <sys/endian.h>
90 #include <sys/systm.h>
91 #include <sys/device.h>
92 #include <sys/sockio.h>
93 #include <sys/mbuf.h>
94 #include <sys/malloc.h>
95 #include <sys/kernel.h>
96 #include <sys/socket.h>
97 
98 #include <net/if.h>
99 #include <net/if_arp.h>
100 #include <net/if_ether.h>
101 #include <net/if_dl.h>
102 #include <net/if_media.h>
103 
104 #include <net/bpf.h>
105 
106 #include <sys/bus.h>
107 
108 #include <dev/mii/mii.h>
109 #include <dev/mii/miivar.h>
110 
111 #include <dev/pci/pcireg.h>
112 #include <dev/pci/pcivar.h>
113 #include <dev/pci/pcidevs.h>
114 
115 #include <dev/pci/if_vgereg.h>
116 
117 #define VGE_IFQ_MAXLEN		64
118 
119 #define VGE_RING_ALIGN		256
120 
121 #define VGE_NTXDESC		256
122 #define VGE_NTXDESC_MASK	(VGE_NTXDESC - 1)
123 #define VGE_NEXT_TXDESC(x)	((x + 1) & VGE_NTXDESC_MASK)
124 #define VGE_PREV_TXDESC(x)	((x - 1) & VGE_NTXDESC_MASK)
125 
126 #define VGE_NRXDESC		256	/* Must be a multiple of 4!! */
127 #define VGE_NRXDESC_MASK	(VGE_NRXDESC - 1)
128 #define VGE_NEXT_RXDESC(x)	((x + 1) & VGE_NRXDESC_MASK)
129 #define VGE_PREV_RXDESC(x)	((x - 1) & VGE_NRXDESC_MASK)
130 
131 #define VGE_ADDR_LO(y)		((uint64_t)(y) & 0xFFFFFFFF)
132 #define VGE_ADDR_HI(y)		((uint64_t)(y) >> 32)
133 #define VGE_BUFLEN(y)		((y) & 0x7FFF)
134 #define ETHER_PAD_LEN		(ETHER_MIN_LEN - ETHER_CRC_LEN)
135 
136 #define VGE_POWER_MANAGEMENT	0	/* disabled for now */
137 
138 /*
139  * Mbuf adjust factor to force 32-bit alignment of IP header.
140  * Drivers should pad ETHER_ALIGN bytes when setting up a
141  * RX mbuf so the upper layers get the IP header properly aligned
142  * past the 14-byte Ethernet header.
143  *
144  * See also comment in vge_encap().
145  */
146 #define ETHER_ALIGN		2
147 
148 #ifdef __NO_STRICT_ALIGNMENT
149 #define VGE_RX_BUFSIZE		MCLBYTES
150 #else
151 #define VGE_RX_PAD		sizeof(uint32_t)
152 #define VGE_RX_BUFSIZE		(MCLBYTES - VGE_RX_PAD)
153 #endif
154 
155 /*
156  * Control structures are DMA'd to the vge chip. We allocate them in
157  * a single clump that maps to a single DMA segment to make several things
158  * easier.
159  */
160 struct vge_control_data {
161 	/* TX descriptors */
162 	struct vge_txdesc	vcd_txdescs[VGE_NTXDESC];
163 	/* RX descriptors */
164 	struct vge_rxdesc	vcd_rxdescs[VGE_NRXDESC];
165 	/* dummy data for TX padding */
166 	uint8_t			vcd_pad[ETHER_PAD_LEN];
167 };
168 
169 #define VGE_CDOFF(x)	offsetof(struct vge_control_data, x)
170 #define VGE_CDTXOFF(x)	VGE_CDOFF(vcd_txdescs[(x)])
171 #define VGE_CDRXOFF(x)	VGE_CDOFF(vcd_rxdescs[(x)])
172 #define VGE_CDPADOFF()	VGE_CDOFF(vcd_pad[0])
173 
174 /*
175  * Software state for TX jobs.
176  */
177 struct vge_txsoft {
178 	struct mbuf	*txs_mbuf;		/* head of our mbuf chain */
179 	bus_dmamap_t	txs_dmamap;		/* our DMA map */
180 };
181 
182 /*
183  * Software state for RX jobs.
184  */
185 struct vge_rxsoft {
186 	struct mbuf	*rxs_mbuf;		/* head of our mbuf chain */
187 	bus_dmamap_t	rxs_dmamap;		/* our DMA map */
188 };
189 
190 
191 struct vge_softc {
192 	device_t		sc_dev;
193 
194 	bus_space_tag_t		sc_bst;		/* bus space tag */
195 	bus_space_handle_t	sc_bsh;		/* bus space handle */
196 	bus_dma_tag_t		sc_dmat;
197 
198 	struct ethercom		sc_ethercom;	/* interface info */
199 	uint8_t			sc_eaddr[ETHER_ADDR_LEN];
200 
201 	void			*sc_intrhand;
202 	struct mii_data		sc_mii;
203 	uint8_t			sc_type;
204 	int			sc_if_flags;
205 	int			sc_link;
206 	int			sc_camidx;
207 	callout_t		sc_timeout;
208 
209 	bus_dmamap_t		sc_cddmamap;
210 #define sc_cddma		sc_cddmamap->dm_segs[0].ds_addr
211 
212 	struct vge_txsoft	sc_txsoft[VGE_NTXDESC];
213 	struct vge_rxsoft	sc_rxsoft[VGE_NRXDESC];
214 	struct vge_control_data	*sc_control_data;
215 #define sc_txdescs		sc_control_data->vcd_txdescs
216 #define sc_rxdescs		sc_control_data->vcd_rxdescs
217 
218 	int			sc_tx_prodidx;
219 	int			sc_tx_considx;
220 	int			sc_tx_free;
221 
222 	struct mbuf		*sc_rx_mhead;
223 	struct mbuf		*sc_rx_mtail;
224 	int			sc_rx_prodidx;
225 	int			sc_rx_consumed;
226 
227 	int			sc_suspended;	/* 0 = normal  1 = suspended */
228 	uint32_t		sc_saved_maps[5];	/* pci data */
229 	uint32_t		sc_saved_biosaddr;
230 	uint8_t			sc_saved_intline;
231 	uint8_t			sc_saved_cachelnsz;
232 	uint8_t			sc_saved_lattimer;
233 };
234 
235 #define VGE_CDTXADDR(sc, x)	((sc)->sc_cddma + VGE_CDTXOFF(x))
236 #define VGE_CDRXADDR(sc, x)	((sc)->sc_cddma + VGE_CDRXOFF(x))
237 #define VGE_CDPADADDR(sc)	((sc)->sc_cddma + VGE_CDPADOFF())
238 
239 #define VGE_TXDESCSYNC(sc, idx, ops)					\
240 	bus_dmamap_sync((sc)->sc_dmat,(sc)->sc_cddmamap,		\
241 	    VGE_CDTXOFF(idx),						\
242 	    offsetof(struct vge_txdesc, td_frag[0]),			\
243 	    (ops))
244 #define VGE_TXFRAGSYNC(sc, idx, nsegs, ops)				\
245 	bus_dmamap_sync((sc)->sc_dmat, (sc)->sc_cddmamap,		\
246 	    VGE_CDTXOFF(idx) +						\
247 	    offsetof(struct vge_txdesc, td_frag[0]),			\
248 	    sizeof(struct vge_txfrag) * (nsegs),			\
249 	    (ops))
250 #define VGE_RXDESCSYNC(sc, idx, ops)					\
251 	bus_dmamap_sync((sc)->sc_dmat, (sc)->sc_cddmamap,		\
252 	    VGE_CDRXOFF(idx),						\
253 	    sizeof(struct vge_rxdesc),					\
254 	    (ops))
255 
256 /*
257  * register space access macros
258  */
259 #define CSR_WRITE_4(sc, reg, val)	\
260 	bus_space_write_4((sc)->sc_bst, (sc)->sc_bsh, (reg), (val))
261 #define CSR_WRITE_2(sc, reg, val)	\
262 	bus_space_write_2((sc)->sc_bst, (sc)->sc_bsh, (reg), (val))
263 #define CSR_WRITE_1(sc, reg, val)	\
264 	bus_space_write_1((sc)->sc_bst, (sc)->sc_bsh, (reg), (val))
265 
266 #define CSR_READ_4(sc, reg)		\
267 	bus_space_read_4((sc)->sc_bst, (sc)->sc_bsh, (reg))
268 #define CSR_READ_2(sc, reg)		\
269 	bus_space_read_2((sc)->sc_bst, (sc)->sc_bsh, (reg))
270 #define CSR_READ_1(sc, reg)		\
271 	bus_space_read_1((sc)->sc_bst, (sc)->sc_bsh, (reg))
272 
273 #define CSR_SETBIT_1(sc, reg, x)	\
274 	CSR_WRITE_1((sc), (reg), CSR_READ_1((sc), (reg)) | (x))
275 #define CSR_SETBIT_2(sc, reg, x)	\
276 	CSR_WRITE_2((sc), (reg), CSR_READ_2((sc), (reg)) | (x))
277 #define CSR_SETBIT_4(sc, reg, x)	\
278 	CSR_WRITE_4((sc), (reg), CSR_READ_4((sc), (reg)) | (x))
279 
280 #define CSR_CLRBIT_1(sc, reg, x)	\
281 	CSR_WRITE_1((sc), (reg), CSR_READ_1((sc), (reg)) & ~(x))
282 #define CSR_CLRBIT_2(sc, reg, x)	\
283 	CSR_WRITE_2((sc), (reg), CSR_READ_2((sc), (reg)) & ~(x))
284 #define CSR_CLRBIT_4(sc, reg, x)	\
285 	CSR_WRITE_4((sc), (reg), CSR_READ_4((sc), (reg)) & ~(x))
286 
287 #define VGE_TIMEOUT		10000
288 
289 #define VGE_PCI_LOIO             0x10
290 #define VGE_PCI_LOMEM            0x14
291 
292 static inline void vge_set_txaddr(struct vge_txfrag *, bus_addr_t);
293 static inline void vge_set_rxaddr(struct vge_rxdesc *, bus_addr_t);
294 
295 static int vge_ifflags_cb(struct ethercom *);
296 
297 static int vge_match(device_t, cfdata_t, void *);
298 static void vge_attach(device_t, device_t, void *);
299 
300 static int vge_encap(struct vge_softc *, struct mbuf *, int);
301 
302 static int vge_allocmem(struct vge_softc *);
303 static int vge_newbuf(struct vge_softc *, int, struct mbuf *);
304 #ifndef __NO_STRICT_ALIGNMENT
305 static inline void vge_fixup_rx(struct mbuf *);
306 #endif
307 static void vge_rxeof(struct vge_softc *);
308 static void vge_txeof(struct vge_softc *);
309 static int vge_intr(void *);
310 static void vge_tick(void *);
311 static void vge_start(struct ifnet *);
312 static int vge_ioctl(struct ifnet *, u_long, void *);
313 static int vge_init(struct ifnet *);
314 static void vge_stop(struct ifnet *, int);
315 static void vge_watchdog(struct ifnet *);
316 #if VGE_POWER_MANAGEMENT
317 static int vge_suspend(device_t);
318 static int vge_resume(device_t);
319 #endif
320 static bool vge_shutdown(device_t, int);
321 
322 static uint16_t vge_read_eeprom(struct vge_softc *, int);
323 
324 static void vge_miipoll_start(struct vge_softc *);
325 static void vge_miipoll_stop(struct vge_softc *);
326 static int vge_miibus_readreg(device_t, int, int);
327 static void vge_miibus_writereg(device_t, int, int, int);
328 static void vge_miibus_statchg(struct ifnet *);
329 
330 static void vge_cam_clear(struct vge_softc *);
331 static int vge_cam_set(struct vge_softc *, uint8_t *);
332 static void vge_setmulti(struct vge_softc *);
333 static void vge_reset(struct vge_softc *);
334 
335 CFATTACH_DECL_NEW(vge, sizeof(struct vge_softc),
336     vge_match, vge_attach, NULL, NULL);
337 
338 static inline void
339 vge_set_txaddr(struct vge_txfrag *f, bus_addr_t daddr)
340 {
341 
342 	f->tf_addrlo = htole32((uint32_t)daddr);
343 	if (sizeof(bus_addr_t) == sizeof(uint64_t))
344 		f->tf_addrhi = htole16(((uint64_t)daddr >> 32) & 0xFFFF);
345 	else
346 		f->tf_addrhi = 0;
347 }
348 
349 static inline void
350 vge_set_rxaddr(struct vge_rxdesc *rxd, bus_addr_t daddr)
351 {
352 
353 	rxd->rd_addrlo = htole32((uint32_t)daddr);
354 	if (sizeof(bus_addr_t) == sizeof(uint64_t))
355 		rxd->rd_addrhi = htole16(((uint64_t)daddr >> 32) & 0xFFFF);
356 	else
357 		rxd->rd_addrhi = 0;
358 }
359 
360 /*
361  * Defragment mbuf chain contents to be as linear as possible.
362  * Returns new mbuf chain on success, NULL on failure. Old mbuf
363  * chain is always freed.
364  * XXX temporary until there would be generic function doing this.
365  */
366 #define m_defrag	vge_m_defrag
367 struct mbuf * vge_m_defrag(struct mbuf *, int);
368 
369 struct mbuf *
370 vge_m_defrag(struct mbuf *mold, int flags)
371 {
372 	struct mbuf *m0, *mn, *n;
373 	size_t sz = mold->m_pkthdr.len;
374 
375 #ifdef DIAGNOSTIC
376 	if ((mold->m_flags & M_PKTHDR) == 0)
377 		panic("m_defrag: not a mbuf chain header");
378 #endif
379 
380 	MGETHDR(m0, flags, MT_DATA);
381 	if (m0 == NULL)
382 		return NULL;
383 	m0->m_pkthdr.len = mold->m_pkthdr.len;
384 	mn = m0;
385 
386 	do {
387 		if (sz > MHLEN) {
388 			MCLGET(mn, M_DONTWAIT);
389 			if ((mn->m_flags & M_EXT) == 0) {
390 				m_freem(m0);
391 				return NULL;
392 			}
393 		}
394 
395 		mn->m_len = MIN(sz, MCLBYTES);
396 
397 		m_copydata(mold, mold->m_pkthdr.len - sz, mn->m_len,
398 		     mtod(mn, void *));
399 
400 		sz -= mn->m_len;
401 
402 		if (sz > 0) {
403 			/* need more mbufs */
404 			MGET(n, M_NOWAIT, MT_DATA);
405 			if (n == NULL) {
406 				m_freem(m0);
407 				return NULL;
408 			}
409 
410 			mn->m_next = n;
411 			mn = n;
412 		}
413 	} while (sz > 0);
414 
415 	return m0;
416 }
417 
418 /*
419  * Read a word of data stored in the EEPROM at address 'addr.'
420  */
421 static uint16_t
422 vge_read_eeprom(struct vge_softc *sc, int addr)
423 {
424 	int i;
425 	uint16_t word = 0;
426 
427 	/*
428 	 * Enter EEPROM embedded programming mode. In order to
429 	 * access the EEPROM at all, we first have to set the
430 	 * EELOAD bit in the CHIPCFG2 register.
431 	 */
432 	CSR_SETBIT_1(sc, VGE_CHIPCFG2, VGE_CHIPCFG2_EELOAD);
433 	CSR_SETBIT_1(sc, VGE_EECSR, VGE_EECSR_EMBP/*|VGE_EECSR_ECS*/);
434 
435 	/* Select the address of the word we want to read */
436 	CSR_WRITE_1(sc, VGE_EEADDR, addr);
437 
438 	/* Issue read command */
439 	CSR_SETBIT_1(sc, VGE_EECMD, VGE_EECMD_ERD);
440 
441 	/* Wait for the done bit to be set. */
442 	for (i = 0; i < VGE_TIMEOUT; i++) {
443 		if (CSR_READ_1(sc, VGE_EECMD) & VGE_EECMD_EDONE)
444 			break;
445 	}
446 
447 	if (i == VGE_TIMEOUT) {
448 		printf("%s: EEPROM read timed out\n", device_xname(sc->sc_dev));
449 		return 0;
450 	}
451 
452 	/* Read the result */
453 	word = CSR_READ_2(sc, VGE_EERDDAT);
454 
455 	/* Turn off EEPROM access mode. */
456 	CSR_CLRBIT_1(sc, VGE_EECSR, VGE_EECSR_EMBP/*|VGE_EECSR_ECS*/);
457 	CSR_CLRBIT_1(sc, VGE_CHIPCFG2, VGE_CHIPCFG2_EELOAD);
458 
459 	return word;
460 }
461 
462 static void
463 vge_miipoll_stop(struct vge_softc *sc)
464 {
465 	int i;
466 
467 	CSR_WRITE_1(sc, VGE_MIICMD, 0);
468 
469 	for (i = 0; i < VGE_TIMEOUT; i++) {
470 		DELAY(1);
471 		if (CSR_READ_1(sc, VGE_MIISTS) & VGE_MIISTS_IIDL)
472 			break;
473 	}
474 
475 	if (i == VGE_TIMEOUT) {
476 		printf("%s: failed to idle MII autopoll\n",
477 		    device_xname(sc->sc_dev));
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 		printf("%s: failed to idle MII autopoll\n",
499 		    device_xname(sc->sc_dev));
500 		return;
501 	}
502 
503 	/* Now enable auto poll mode. */
504 
505 	CSR_WRITE_1(sc, VGE_MIICMD, VGE_MIICMD_MAUTO);
506 
507 	/* And make sure it started. */
508 
509 	for (i = 0; i < VGE_TIMEOUT; i++) {
510 		DELAY(1);
511 		if ((CSR_READ_1(sc, VGE_MIISTS) & VGE_MIISTS_IIDL) == 0)
512 			break;
513 	}
514 
515 	if (i == VGE_TIMEOUT) {
516 		printf("%s: failed to start MII autopoll\n",
517 		    device_xname(sc->sc_dev));
518 	}
519 }
520 
521 static int
522 vge_miibus_readreg(device_t dev, int phy, int reg)
523 {
524 	struct vge_softc *sc;
525 	int i, s;
526 	uint16_t rval;
527 
528 	sc = device_private(dev);
529 	rval = 0;
530 	if (phy != (CSR_READ_1(sc, VGE_MIICFG) & 0x1F))
531 		return 0;
532 
533 	s = splnet();
534 	vge_miipoll_stop(sc);
535 
536 	/* Specify the register we want to read. */
537 	CSR_WRITE_1(sc, VGE_MIIADDR, reg);
538 
539 	/* Issue read command. */
540 	CSR_SETBIT_1(sc, VGE_MIICMD, VGE_MIICMD_RCMD);
541 
542 	/* Wait for the read command bit to self-clear. */
543 	for (i = 0; i < VGE_TIMEOUT; i++) {
544 		DELAY(1);
545 		if ((CSR_READ_1(sc, VGE_MIICMD) & VGE_MIICMD_RCMD) == 0)
546 			break;
547 	}
548 
549 	if (i == VGE_TIMEOUT)
550 		printf("%s: MII read timed out\n", device_xname(sc->sc_dev));
551 	else
552 		rval = CSR_READ_2(sc, VGE_MIIDATA);
553 
554 	vge_miipoll_start(sc);
555 	splx(s);
556 
557 	return rval;
558 }
559 
560 static void
561 vge_miibus_writereg(device_t dev, int phy, int reg, int data)
562 {
563 	struct vge_softc *sc;
564 	int i, s;
565 
566 	sc = device_private(dev);
567 	if (phy != (CSR_READ_1(sc, VGE_MIICFG) & 0x1F))
568 		return;
569 
570 	s = splnet();
571 	vge_miipoll_stop(sc);
572 
573 	/* Specify the register we want to write. */
574 	CSR_WRITE_1(sc, VGE_MIIADDR, reg);
575 
576 	/* Specify the data we want to write. */
577 	CSR_WRITE_2(sc, VGE_MIIDATA, data);
578 
579 	/* Issue write command. */
580 	CSR_SETBIT_1(sc, VGE_MIICMD, VGE_MIICMD_WCMD);
581 
582 	/* Wait for the write command bit to self-clear. */
583 	for (i = 0; i < VGE_TIMEOUT; i++) {
584 		DELAY(1);
585 		if ((CSR_READ_1(sc, VGE_MIICMD) & VGE_MIICMD_WCMD) == 0)
586 			break;
587 	}
588 
589 	if (i == VGE_TIMEOUT) {
590 		printf("%s: MII write timed out\n", device_xname(sc->sc_dev));
591 	}
592 
593 	vge_miipoll_start(sc);
594 	splx(s);
595 }
596 
597 static void
598 vge_cam_clear(struct vge_softc *sc)
599 {
600 	int i;
601 
602 	/*
603 	 * Turn off all the mask bits. This tells the chip
604 	 * that none of the entries in the CAM filter are valid.
605 	 * desired entries will be enabled as we fill the filter in.
606 	 */
607 
608 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
609 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_CAMMASK);
610 	CSR_WRITE_1(sc, VGE_CAMADDR, VGE_CAMADDR_ENABLE);
611 	for (i = 0; i < 8; i++)
612 		CSR_WRITE_1(sc, VGE_CAM0 + i, 0);
613 
614 	/* Clear the VLAN filter too. */
615 
616 	CSR_WRITE_1(sc, VGE_CAMADDR, VGE_CAMADDR_ENABLE|VGE_CAMADDR_AVSEL|0);
617 	for (i = 0; i < 8; i++)
618 		CSR_WRITE_1(sc, VGE_CAM0 + i, 0);
619 
620 	CSR_WRITE_1(sc, VGE_CAMADDR, 0);
621 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
622 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_MAR);
623 
624 	sc->sc_camidx = 0;
625 }
626 
627 static int
628 vge_cam_set(struct vge_softc *sc, uint8_t *addr)
629 {
630 	int i, error;
631 
632 	error = 0;
633 
634 	if (sc->sc_camidx == VGE_CAM_MAXADDRS)
635 		return ENOSPC;
636 
637 	/* Select the CAM data page. */
638 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
639 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_CAMDATA);
640 
641 	/* Set the filter entry we want to update and enable writing. */
642 	CSR_WRITE_1(sc, VGE_CAMADDR, VGE_CAMADDR_ENABLE | sc->sc_camidx);
643 
644 	/* Write the address to the CAM registers */
645 	for (i = 0; i < ETHER_ADDR_LEN; i++)
646 		CSR_WRITE_1(sc, VGE_CAM0 + i, addr[i]);
647 
648 	/* Issue a write command. */
649 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_WRITE);
650 
651 	/* Wake for it to clear. */
652 	for (i = 0; i < VGE_TIMEOUT; i++) {
653 		DELAY(1);
654 		if ((CSR_READ_1(sc, VGE_CAMCTL) & VGE_CAMCTL_WRITE) == 0)
655 			break;
656 	}
657 
658 	if (i == VGE_TIMEOUT) {
659 		printf("%s: setting CAM filter failed\n",
660 		    device_xname(sc->sc_dev));
661 		error = EIO;
662 		goto fail;
663 	}
664 
665 	/* Select the CAM mask page. */
666 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
667 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_CAMMASK);
668 
669 	/* Set the mask bit that enables this filter. */
670 	CSR_SETBIT_1(sc, VGE_CAM0 + (sc->sc_camidx / 8),
671 	    1 << (sc->sc_camidx & 7));
672 
673 	sc->sc_camidx++;
674 
675  fail:
676 	/* Turn off access to CAM. */
677 	CSR_WRITE_1(sc, VGE_CAMADDR, 0);
678 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
679 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_MAR);
680 
681 	return error;
682 }
683 
684 /*
685  * Program the multicast filter. We use the 64-entry CAM filter
686  * for perfect filtering. If there's more than 64 multicast addresses,
687  * we use the hash filter instead.
688  */
689 static void
690 vge_setmulti(struct vge_softc *sc)
691 {
692 	struct ifnet *ifp;
693 	int error;
694 	uint32_t h, hashes[2] = { 0, 0 };
695 	struct ether_multi *enm;
696 	struct ether_multistep step;
697 
698 	error = 0;
699 	ifp = &sc->sc_ethercom.ec_if;
700 
701 	/* First, zot all the multicast entries. */
702 	vge_cam_clear(sc);
703 	CSR_WRITE_4(sc, VGE_MAR0, 0);
704 	CSR_WRITE_4(sc, VGE_MAR1, 0);
705 	ifp->if_flags &= ~IFF_ALLMULTI;
706 
707 	/*
708 	 * If the user wants allmulti or promisc mode, enable reception
709 	 * of all multicast frames.
710 	 */
711 	if (ifp->if_flags & IFF_PROMISC) {
712  allmulti:
713 		CSR_WRITE_4(sc, VGE_MAR0, 0xFFFFFFFF);
714 		CSR_WRITE_4(sc, VGE_MAR1, 0xFFFFFFFF);
715 		ifp->if_flags |= IFF_ALLMULTI;
716 		return;
717 	}
718 
719 	/* Now program new ones */
720 	ETHER_FIRST_MULTI(step, &sc->sc_ethercom, enm);
721 	while (enm != NULL) {
722 		/*
723 		 * If multicast range, fall back to ALLMULTI.
724 		 */
725 		if (memcmp(enm->enm_addrlo, enm->enm_addrhi,
726 				ETHER_ADDR_LEN) != 0)
727 			goto allmulti;
728 
729 		error = vge_cam_set(sc, enm->enm_addrlo);
730 		if (error)
731 			break;
732 
733 		ETHER_NEXT_MULTI(step, enm);
734 	}
735 
736 	/* If there were too many addresses, use the hash filter. */
737 	if (error) {
738 		vge_cam_clear(sc);
739 
740 		ETHER_FIRST_MULTI(step, &sc->sc_ethercom, enm);
741 		while (enm != NULL) {
742 			/*
743 			 * If multicast range, fall back to ALLMULTI.
744 			 */
745 			if (memcmp(enm->enm_addrlo, enm->enm_addrhi,
746 					ETHER_ADDR_LEN) != 0)
747 				goto allmulti;
748 
749 			h = ether_crc32_be(enm->enm_addrlo,
750 			    ETHER_ADDR_LEN) >> 26;
751 			hashes[h >> 5] |= 1 << (h & 0x1f);
752 
753 			ETHER_NEXT_MULTI(step, enm);
754 		}
755 
756 		CSR_WRITE_4(sc, VGE_MAR0, hashes[0]);
757 		CSR_WRITE_4(sc, VGE_MAR1, hashes[1]);
758 	}
759 }
760 
761 static void
762 vge_reset(struct vge_softc *sc)
763 {
764 	int i;
765 
766 	CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_SOFTRESET);
767 
768 	for (i = 0; i < VGE_TIMEOUT; i++) {
769 		DELAY(5);
770 		if ((CSR_READ_1(sc, VGE_CRS1) & VGE_CR1_SOFTRESET) == 0)
771 			break;
772 	}
773 
774 	if (i == VGE_TIMEOUT) {
775 		printf("%s: soft reset timed out", device_xname(sc->sc_dev));
776 		CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_STOP_FORCE);
777 		DELAY(2000);
778 	}
779 
780 	DELAY(5000);
781 
782 	CSR_SETBIT_1(sc, VGE_EECSR, VGE_EECSR_RELOAD);
783 
784 	for (i = 0; i < VGE_TIMEOUT; i++) {
785 		DELAY(5);
786 		if ((CSR_READ_1(sc, VGE_EECSR) & VGE_EECSR_RELOAD) == 0)
787 			break;
788 	}
789 
790 	if (i == VGE_TIMEOUT) {
791 		printf("%s: EEPROM reload timed out\n",
792 		    device_xname(sc->sc_dev));
793 		return;
794 	}
795 
796 	/*
797 	 * On some machine, the first read data from EEPROM could be
798 	 * messed up, so read one dummy data here to avoid the mess.
799 	 */
800 	(void)vge_read_eeprom(sc, 0);
801 
802 	CSR_CLRBIT_1(sc, VGE_CHIPCFG0, VGE_CHIPCFG0_PACPI);
803 }
804 
805 /*
806  * Probe for a VIA gigabit chip. Check the PCI vendor and device
807  * IDs against our list and return a device name if we find a match.
808  */
809 static int
810 vge_match(device_t parent, cfdata_t match, void *aux)
811 {
812 	struct pci_attach_args *pa = aux;
813 
814 	if (PCI_VENDOR(pa->pa_id) == PCI_VENDOR_VIATECH
815 	    && PCI_PRODUCT(pa->pa_id) == PCI_PRODUCT_VIATECH_VT612X)
816 		return 1;
817 
818 	return 0;
819 }
820 
821 static int
822 vge_allocmem(struct vge_softc *sc)
823 {
824 	int error;
825 	int nseg;
826 	int i;
827 	bus_dma_segment_t seg;
828 
829 	/*
830 	 * Allocate memory for control data.
831 	 */
832 
833 	error = bus_dmamem_alloc(sc->sc_dmat, sizeof(struct vge_control_data),
834 	     VGE_RING_ALIGN, 0, &seg, 1, &nseg, BUS_DMA_NOWAIT);
835 	if (error) {
836 		aprint_error_dev(sc->sc_dev,
837 		    "could not allocate control data dma memory\n");
838 		goto fail_1;
839 	}
840 
841 	/* Map the memory to kernel VA space */
842 
843 	error = bus_dmamem_map(sc->sc_dmat, &seg, nseg,
844 	    sizeof(struct vge_control_data), (void **)&sc->sc_control_data,
845 	    BUS_DMA_NOWAIT);
846 	if (error) {
847 		aprint_error_dev(sc->sc_dev,
848 		    "could not map control data dma memory\n");
849 		goto fail_2;
850 	}
851 	memset(sc->sc_control_data, 0, sizeof(struct vge_control_data));
852 
853 	/*
854 	 * Create map for control data.
855 	 */
856 	error = bus_dmamap_create(sc->sc_dmat,
857 	    sizeof(struct vge_control_data), 1,
858 	    sizeof(struct vge_control_data), 0, BUS_DMA_NOWAIT,
859 	    &sc->sc_cddmamap);
860 	if (error) {
861 		aprint_error_dev(sc->sc_dev,
862 		    "could not create control data dmamap\n");
863 		goto fail_3;
864 	}
865 
866 	/* Load the map for the control data. */
867 	error = bus_dmamap_load(sc->sc_dmat, sc->sc_cddmamap,
868 	    sc->sc_control_data, sizeof(struct vge_control_data), NULL,
869 	    BUS_DMA_NOWAIT);
870 	if (error) {
871 		aprint_error_dev(sc->sc_dev,
872 		    "could not load control data dma memory\n");
873 		goto fail_4;
874 	}
875 
876 	/* Create DMA maps for TX buffers */
877 
878 	for (i = 0; i < VGE_NTXDESC; i++) {
879 		error = bus_dmamap_create(sc->sc_dmat, VGE_TX_MAXLEN,
880 		    VGE_TX_FRAGS, VGE_TX_MAXLEN, 0, BUS_DMA_NOWAIT,
881 		    &sc->sc_txsoft[i].txs_dmamap);
882 		if (error) {
883 			aprint_error_dev(sc->sc_dev,
884 			    "can't create DMA map for TX descs\n");
885 			goto fail_5;
886 		}
887 	}
888 
889 	/* Create DMA maps for RX buffers */
890 
891 	for (i = 0; i < VGE_NRXDESC; i++) {
892 		error = bus_dmamap_create(sc->sc_dmat, MCLBYTES,
893 		    1, MCLBYTES, 0, BUS_DMA_NOWAIT,
894 		    &sc->sc_rxsoft[i].rxs_dmamap);
895 		if (error) {
896 			aprint_error_dev(sc->sc_dev,
897 			    "can't create DMA map for RX descs\n");
898 			goto fail_6;
899 		}
900 		sc->sc_rxsoft[i].rxs_mbuf = NULL;
901 	}
902 
903 	return 0;
904 
905  fail_6:
906 	for (i = 0; i < VGE_NRXDESC; i++) {
907 		if (sc->sc_rxsoft[i].rxs_dmamap != NULL)
908 			bus_dmamap_destroy(sc->sc_dmat,
909 			    sc->sc_rxsoft[i].rxs_dmamap);
910 	}
911  fail_5:
912 	for (i = 0; i < VGE_NTXDESC; i++) {
913 		if (sc->sc_txsoft[i].txs_dmamap != NULL)
914 			bus_dmamap_destroy(sc->sc_dmat,
915 			    sc->sc_txsoft[i].txs_dmamap);
916 	}
917 	bus_dmamap_unload(sc->sc_dmat, sc->sc_cddmamap);
918  fail_4:
919 	bus_dmamap_destroy(sc->sc_dmat, sc->sc_cddmamap);
920  fail_3:
921 	bus_dmamem_unmap(sc->sc_dmat, (void *)sc->sc_control_data,
922 	    sizeof(struct vge_control_data));
923  fail_2:
924 	bus_dmamem_free(sc->sc_dmat, &seg, nseg);
925  fail_1:
926 	return ENOMEM;
927 }
928 
929 /*
930  * Attach the interface. Allocate softc structures, do ifmedia
931  * setup and ethernet/BPF attach.
932  */
933 static void
934 vge_attach(device_t parent, device_t self, void *aux)
935 {
936 	uint8_t	*eaddr;
937 	struct vge_softc *sc = device_private(self);
938 	struct ifnet *ifp;
939 	struct pci_attach_args *pa = aux;
940 	pci_chipset_tag_t pc = pa->pa_pc;
941 	const char *intrstr;
942 	pci_intr_handle_t ih;
943 	uint16_t val;
944 	char intrbuf[PCI_INTRSTR_LEN];
945 
946 	sc->sc_dev = self;
947 
948 	pci_aprint_devinfo_fancy(pa, NULL, "VIA VT612X Gigabit Ethernet", 1);
949 
950 	/* Make sure bus-mastering is enabled */
951         pci_conf_write(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG,
952 	    pci_conf_read(pc, pa->pa_tag, PCI_COMMAND_STATUS_REG) |
953 	    PCI_COMMAND_MASTER_ENABLE);
954 
955 	/*
956 	 * Map control/status registers.
957 	 */
958 	if (pci_mapreg_map(pa, VGE_PCI_LOMEM, PCI_MAPREG_TYPE_MEM, 0,
959 	    &sc->sc_bst, &sc->sc_bsh, NULL, NULL) != 0) {
960 		aprint_error_dev(self, "couldn't map memory\n");
961 		return;
962 	}
963 
964         /*
965          * Map and establish our interrupt.
966          */
967 	if (pci_intr_map(pa, &ih)) {
968 		aprint_error_dev(self, "unable to map interrupt\n");
969 		return;
970 	}
971 	intrstr = pci_intr_string(pc, ih, intrbuf, sizeof(intrbuf));
972 	sc->sc_intrhand = pci_intr_establish(pc, ih, IPL_NET, vge_intr, sc);
973 	if (sc->sc_intrhand == NULL) {
974 		aprint_error_dev(self, "unable to establish interrupt");
975 		if (intrstr != NULL)
976 			aprint_error(" at %s", intrstr);
977 		aprint_error("\n");
978 		return;
979 	}
980 	aprint_normal_dev(self, "interrupting at %s\n", intrstr);
981 
982 	/* Reset the adapter. */
983 	vge_reset(sc);
984 
985 	/*
986 	 * Get station address from the EEPROM.
987 	 */
988 	eaddr = sc->sc_eaddr;
989 	val = vge_read_eeprom(sc, VGE_EE_EADDR + 0);
990 	eaddr[0] = val & 0xff;
991 	eaddr[1] = val >> 8;
992 	val = vge_read_eeprom(sc, VGE_EE_EADDR + 1);
993 	eaddr[2] = val & 0xff;
994 	eaddr[3] = val >> 8;
995 	val = vge_read_eeprom(sc, VGE_EE_EADDR + 2);
996 	eaddr[4] = val & 0xff;
997 	eaddr[5] = val >> 8;
998 
999 	aprint_normal_dev(self, "Ethernet address: %s\n",
1000 	    ether_sprintf(eaddr));
1001 
1002 	/*
1003 	 * Use the 32bit tag. Hardware supports 48bit physical addresses,
1004 	 * but we don't use that for now.
1005 	 */
1006 	sc->sc_dmat = pa->pa_dmat;
1007 
1008 	if (vge_allocmem(sc) != 0)
1009 		return;
1010 
1011 	ifp = &sc->sc_ethercom.ec_if;
1012 	ifp->if_softc = sc;
1013 	strlcpy(ifp->if_xname, device_xname(self), IFNAMSIZ);
1014 	ifp->if_mtu = ETHERMTU;
1015 	ifp->if_baudrate = IF_Gbps(1);
1016 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
1017 	ifp->if_ioctl = vge_ioctl;
1018 	ifp->if_start = vge_start;
1019 	ifp->if_init = vge_init;
1020 	ifp->if_stop = vge_stop;
1021 
1022 	/*
1023 	 * We can support 802.1Q VLAN-sized frames and jumbo
1024 	 * Ethernet frames.
1025 	 */
1026 	sc->sc_ethercom.ec_capabilities |=
1027 	    ETHERCAP_VLAN_MTU | ETHERCAP_JUMBO_MTU |
1028 	    ETHERCAP_VLAN_HWTAGGING;
1029 
1030 	/*
1031 	 * We can do IPv4/TCPv4/UDPv4 checksums in hardware.
1032 	 */
1033 	ifp->if_capabilities |=
1034 	    IFCAP_CSUM_IPv4_Tx | IFCAP_CSUM_IPv4_Rx |
1035 	    IFCAP_CSUM_TCPv4_Tx | IFCAP_CSUM_TCPv4_Rx |
1036 	    IFCAP_CSUM_UDPv4_Tx | IFCAP_CSUM_UDPv4_Rx;
1037 
1038 #ifdef DEVICE_POLLING
1039 #ifdef IFCAP_POLLING
1040 	ifp->if_capabilities |= IFCAP_POLLING;
1041 #endif
1042 #endif
1043 	ifp->if_watchdog = vge_watchdog;
1044 	IFQ_SET_MAXLEN(&ifp->if_snd, max(VGE_IFQ_MAXLEN, IFQ_MAXLEN));
1045 	IFQ_SET_READY(&ifp->if_snd);
1046 
1047 	/*
1048 	 * Initialize our media structures and probe the MII.
1049 	 */
1050 	sc->sc_mii.mii_ifp = ifp;
1051 	sc->sc_mii.mii_readreg = vge_miibus_readreg;
1052 	sc->sc_mii.mii_writereg = vge_miibus_writereg;
1053 	sc->sc_mii.mii_statchg = vge_miibus_statchg;
1054 
1055 	sc->sc_ethercom.ec_mii = &sc->sc_mii;
1056 	ifmedia_init(&sc->sc_mii.mii_media, 0, ether_mediachange,
1057 	    ether_mediastatus);
1058 	mii_attach(self, &sc->sc_mii, 0xffffffff, MII_PHY_ANY,
1059 	    MII_OFFSET_ANY, MIIF_DOPAUSE);
1060 	if (LIST_FIRST(&sc->sc_mii.mii_phys) == NULL) {
1061 		ifmedia_add(&sc->sc_mii.mii_media, IFM_ETHER|IFM_NONE, 0, NULL);
1062 		ifmedia_set(&sc->sc_mii.mii_media, IFM_ETHER|IFM_NONE);
1063 	} else
1064 		ifmedia_set(&sc->sc_mii.mii_media, IFM_ETHER|IFM_AUTO);
1065 
1066 	/*
1067 	 * Attach the interface.
1068 	 */
1069 	if_attach(ifp);
1070 	if_deferred_start_init(ifp, NULL);
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 		m_set_rcvif(m, ifp);
1331 
1332 		/* Do RX checksumming if enabled */
1333 		if (ifp->if_csum_flags_rx & M_CSUM_IPv4) {
1334 
1335 			/* Check IP header checksum */
1336 			if (rxctl & VGE_RDCTL_IPPKT)
1337 				m->m_pkthdr.csum_flags |= M_CSUM_IPv4;
1338 			if ((rxctl & VGE_RDCTL_IPCSUMOK) == 0)
1339 				m->m_pkthdr.csum_flags |= M_CSUM_IPv4_BAD;
1340 		}
1341 
1342 		if (ifp->if_csum_flags_rx & M_CSUM_TCPv4) {
1343 			/* Check UDP checksum */
1344 			if (rxctl & VGE_RDCTL_TCPPKT)
1345 				m->m_pkthdr.csum_flags |= M_CSUM_TCPv4;
1346 
1347 			if ((rxctl & VGE_RDCTL_PROTOCSUMOK) == 0)
1348 				m->m_pkthdr.csum_flags |= M_CSUM_TCP_UDP_BAD;
1349 		}
1350 
1351 		if (ifp->if_csum_flags_rx & M_CSUM_UDPv4) {
1352 			/* Check UDP checksum */
1353 			if (rxctl & VGE_RDCTL_UDPPKT)
1354 				m->m_pkthdr.csum_flags |= M_CSUM_UDPv4;
1355 
1356 			if ((rxctl & VGE_RDCTL_PROTOCSUMOK) == 0)
1357 				m->m_pkthdr.csum_flags |= M_CSUM_TCP_UDP_BAD;
1358 		}
1359 
1360 		if (rxstat & VGE_RDSTS_VTAG) {
1361 			/*
1362 			 * We use bswap16() here because:
1363 			 * On LE machines, tag is stored in BE as stream data.
1364 			 * On BE machines, tag is stored in BE as stream data
1365 			 *  but it was already swapped by le32toh() above.
1366 			 */
1367 			VLAN_INPUT_TAG(ifp, m,
1368 			    bswap16(rxctl & VGE_RDCTL_VLANID), continue);
1369 		}
1370 
1371 		if_percpuq_enqueue(ifp->if_percpuq, m);
1372 
1373 		lim++;
1374 		if (lim == VGE_NRXDESC)
1375 			break;
1376 	}
1377 
1378 	sc->sc_rx_prodidx = idx;
1379 	CSR_WRITE_2(sc, VGE_RXDESC_RESIDUECNT, lim);
1380 }
1381 
1382 static void
1383 vge_txeof(struct vge_softc *sc)
1384 {
1385 	struct ifnet *ifp;
1386 	struct vge_txsoft *txs;
1387 	uint32_t txstat;
1388 	int idx;
1389 
1390 	ifp = &sc->sc_ethercom.ec_if;
1391 
1392 	for (idx = sc->sc_tx_considx;
1393 	    sc->sc_tx_free < VGE_NTXDESC;
1394 	    idx = VGE_NEXT_TXDESC(idx), sc->sc_tx_free++) {
1395 		VGE_TXDESCSYNC(sc, idx,
1396 		    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1397 		txstat = le32toh(sc->sc_txdescs[idx].td_sts);
1398 		VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1399 		if (txstat & VGE_TDSTS_OWN) {
1400 			break;
1401 		}
1402 
1403 		txs = &sc->sc_txsoft[idx];
1404 		m_freem(txs->txs_mbuf);
1405 		txs->txs_mbuf = NULL;
1406 		bus_dmamap_sync(sc->sc_dmat, txs->txs_dmamap, 0,
1407 		    txs->txs_dmamap->dm_mapsize, BUS_DMASYNC_POSTWRITE);
1408 		bus_dmamap_unload(sc->sc_dmat, txs->txs_dmamap);
1409 		if (txstat & (VGE_TDSTS_EXCESSCOLL|VGE_TDSTS_COLL))
1410 			ifp->if_collisions++;
1411 		if (txstat & VGE_TDSTS_TXERR)
1412 			ifp->if_oerrors++;
1413 		else
1414 			ifp->if_opackets++;
1415 	}
1416 
1417 	sc->sc_tx_considx = idx;
1418 
1419 	if (sc->sc_tx_free > 0) {
1420 		ifp->if_flags &= ~IFF_OACTIVE;
1421 	}
1422 
1423 	/*
1424 	 * If not all descriptors have been released reaped yet,
1425 	 * reload the timer so that we will eventually get another
1426 	 * interrupt that will cause us to re-enter this routine.
1427 	 * This is done in case the transmitter has gone idle.
1428 	 */
1429 	if (sc->sc_tx_free < VGE_NTXDESC)
1430 		CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_TIMER0_ENABLE);
1431 	else
1432 		ifp->if_timer = 0;
1433 }
1434 
1435 static void
1436 vge_tick(void *arg)
1437 {
1438 	struct vge_softc *sc;
1439 	struct ifnet *ifp;
1440 	struct mii_data *mii;
1441 	int s;
1442 
1443 	sc = arg;
1444 	ifp = &sc->sc_ethercom.ec_if;
1445 	mii = &sc->sc_mii;
1446 
1447 	s = splnet();
1448 
1449 	callout_schedule(&sc->sc_timeout, hz);
1450 
1451 	mii_tick(mii);
1452 	if (sc->sc_link) {
1453 		if ((mii->mii_media_status & IFM_ACTIVE) == 0)
1454 			sc->sc_link = 0;
1455 	} else {
1456 		if (mii->mii_media_status & IFM_ACTIVE &&
1457 		    IFM_SUBTYPE(mii->mii_media_active) != IFM_NONE) {
1458 			sc->sc_link = 1;
1459 			if (!IFQ_IS_EMPTY(&ifp->if_snd))
1460 				vge_start(ifp);
1461 		}
1462 	}
1463 
1464 	splx(s);
1465 }
1466 
1467 static int
1468 vge_intr(void *arg)
1469 {
1470 	struct vge_softc *sc;
1471 	struct ifnet *ifp;
1472 	uint32_t status;
1473 	int claim;
1474 
1475 	sc = arg;
1476 	claim = 0;
1477 	if (sc->sc_suspended) {
1478 		return claim;
1479 	}
1480 
1481 	ifp = &sc->sc_ethercom.ec_if;
1482 
1483 	if ((ifp->if_flags & IFF_UP) == 0) {
1484 		return claim;
1485 	}
1486 
1487 	/* Disable interrupts */
1488 	CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
1489 
1490 	for (;;) {
1491 
1492 		status = CSR_READ_4(sc, VGE_ISR);
1493 		/* If the card has gone away the read returns 0xffffffff. */
1494 		if (status == 0xFFFFFFFF)
1495 			break;
1496 
1497 		if (status) {
1498 			claim = 1;
1499 			CSR_WRITE_4(sc, VGE_ISR, status);
1500 		}
1501 
1502 		if ((status & VGE_INTRS) == 0)
1503 			break;
1504 
1505 		if (status & (VGE_ISR_RXOK|VGE_ISR_RXOK_HIPRIO))
1506 			vge_rxeof(sc);
1507 
1508 		if (status & (VGE_ISR_RXOFLOW|VGE_ISR_RXNODESC)) {
1509 			vge_rxeof(sc);
1510 			CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_RUN);
1511 			CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_WAK);
1512 		}
1513 
1514 		if (status & (VGE_ISR_TXOK0|VGE_ISR_TIMER0))
1515 			vge_txeof(sc);
1516 
1517 		if (status & (VGE_ISR_TXDMA_STALL|VGE_ISR_RXDMA_STALL))
1518 			vge_init(ifp);
1519 
1520 		if (status & VGE_ISR_LINKSTS)
1521 			vge_tick(sc);
1522 	}
1523 
1524 	/* Re-enable interrupts */
1525 	CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_GMSK);
1526 
1527 	if (claim)
1528 		if_schedule_deferred_start(ifp);
1529 
1530 	return claim;
1531 }
1532 
1533 static int
1534 vge_encap(struct vge_softc *sc, struct mbuf *m_head, int idx)
1535 {
1536 	struct vge_txsoft *txs;
1537 	struct vge_txdesc *txd;
1538 	struct vge_txfrag *f;
1539 	struct mbuf *m_new;
1540 	bus_dmamap_t map;
1541 	int m_csumflags, seg, error, flags;
1542 	struct m_tag *mtag;
1543 	size_t sz;
1544 	uint32_t td_sts, td_ctl;
1545 
1546 	KASSERT(sc->sc_tx_free > 0);
1547 
1548 	txd = &sc->sc_txdescs[idx];
1549 
1550 #ifdef DIAGNOSTIC
1551 	/* If this descriptor is still owned by the chip, bail. */
1552 	VGE_TXDESCSYNC(sc, idx,
1553 	    BUS_DMASYNC_POSTREAD|BUS_DMASYNC_POSTWRITE);
1554 	td_sts = le32toh(txd->td_sts);
1555 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD);
1556 	if (td_sts & VGE_TDSTS_OWN) {
1557 		return ENOBUFS;
1558 	}
1559 #endif
1560 
1561 	/*
1562 	 * Preserve m_pkthdr.csum_flags here since m_head might be
1563 	 * updated by m_defrag()
1564 	 */
1565 	m_csumflags = m_head->m_pkthdr.csum_flags;
1566 
1567 	txs = &sc->sc_txsoft[idx];
1568 	map = txs->txs_dmamap;
1569 	error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m_head, BUS_DMA_NOWAIT);
1570 
1571 	/* If too many segments to map, coalesce */
1572 	if (error == EFBIG ||
1573 	    (m_head->m_pkthdr.len < ETHER_PAD_LEN &&
1574 	     map->dm_nsegs == VGE_TX_FRAGS)) {
1575 		m_new = m_defrag(m_head, M_DONTWAIT);
1576 		if (m_new == NULL)
1577 			return EFBIG;
1578 
1579 		error = bus_dmamap_load_mbuf(sc->sc_dmat, map,
1580 		    m_new, BUS_DMA_NOWAIT);
1581 		if (error) {
1582 			m_freem(m_new);
1583 			return error;
1584 		}
1585 
1586 		m_head = m_new;
1587 	} else if (error)
1588 		return error;
1589 
1590 	txs->txs_mbuf = m_head;
1591 
1592 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1593 	    BUS_DMASYNC_PREWRITE);
1594 
1595 	for (seg = 0, f = &txd->td_frag[0]; seg < map->dm_nsegs; seg++, f++) {
1596 		f->tf_buflen = htole16(VGE_BUFLEN(map->dm_segs[seg].ds_len));
1597 		vge_set_txaddr(f, map->dm_segs[seg].ds_addr);
1598 	}
1599 
1600 	/* Argh. This chip does not autopad short frames */
1601 	sz = m_head->m_pkthdr.len;
1602 	if (sz < ETHER_PAD_LEN) {
1603 		f->tf_buflen = htole16(VGE_BUFLEN(ETHER_PAD_LEN - sz));
1604 		vge_set_txaddr(f, VGE_CDPADADDR(sc));
1605 		sz = ETHER_PAD_LEN;
1606 		seg++;
1607 	}
1608 	VGE_TXFRAGSYNC(sc, idx, seg, BUS_DMASYNC_PREWRITE);
1609 
1610 	/*
1611 	 * When telling the chip how many segments there are, we
1612 	 * must use nsegs + 1 instead of just nsegs. Darned if I
1613 	 * know why.
1614 	 */
1615 	seg++;
1616 
1617 	flags = 0;
1618 	if (m_csumflags & M_CSUM_IPv4)
1619 		flags |= VGE_TDCTL_IPCSUM;
1620 	if (m_csumflags & M_CSUM_TCPv4)
1621 		flags |= VGE_TDCTL_TCPCSUM;
1622 	if (m_csumflags & M_CSUM_UDPv4)
1623 		flags |= VGE_TDCTL_UDPCSUM;
1624 	td_sts = sz << 16;
1625 	td_ctl = flags | (seg << 28) | VGE_TD_LS_NORM;
1626 
1627 	if (sz > ETHERMTU + ETHER_HDR_LEN)
1628 		td_ctl |= VGE_TDCTL_JUMBO;
1629 
1630 	/*
1631 	 * Set up hardware VLAN tagging.
1632 	 */
1633 	mtag = VLAN_OUTPUT_TAG(&sc->sc_ethercom, m_head);
1634 	if (mtag != NULL) {
1635 		/*
1636 		 * No need htons() here since vge(4) chip assumes
1637 		 * that tags are written in little endian and
1638 		 * we already use htole32() here.
1639 		 */
1640 		td_ctl |= VLAN_TAG_VALUE(mtag) | VGE_TDCTL_VTAG;
1641 	}
1642 	txd->td_ctl = htole32(td_ctl);
1643 	txd->td_sts = htole32(td_sts);
1644 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1645 
1646 	txd->td_sts = htole32(VGE_TDSTS_OWN | td_sts);
1647 	VGE_TXDESCSYNC(sc, idx, BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1648 
1649 	sc->sc_tx_free--;
1650 
1651 	return 0;
1652 }
1653 
1654 /*
1655  * Main transmit routine.
1656  */
1657 
1658 static void
1659 vge_start(struct ifnet *ifp)
1660 {
1661 	struct vge_softc *sc;
1662 	struct vge_txsoft *txs;
1663 	struct mbuf *m_head;
1664 	int idx, pidx, ofree, error;
1665 
1666 	sc = ifp->if_softc;
1667 
1668 	if (!sc->sc_link ||
1669 	    (ifp->if_flags & (IFF_RUNNING|IFF_OACTIVE)) != IFF_RUNNING) {
1670 		return;
1671 	}
1672 
1673 	m_head = NULL;
1674 	idx = sc->sc_tx_prodidx;
1675 	pidx = VGE_PREV_TXDESC(idx);
1676 	ofree = sc->sc_tx_free;
1677 
1678 	/*
1679 	 * Loop through the send queue, setting up transmit descriptors
1680 	 * until we drain the queue, or use up all available transmit
1681 	 * descriptors.
1682 	 */
1683 	for (;;) {
1684 		/* Grab a packet off the queue. */
1685 		IFQ_POLL(&ifp->if_snd, m_head);
1686 		if (m_head == NULL)
1687 			break;
1688 
1689 		if (sc->sc_tx_free == 0) {
1690 			/*
1691 			 * All slots used, stop for now.
1692 			 */
1693 			ifp->if_flags |= IFF_OACTIVE;
1694 			break;
1695 		}
1696 
1697 		txs = &sc->sc_txsoft[idx];
1698 		KASSERT(txs->txs_mbuf == NULL);
1699 
1700 		if ((error = vge_encap(sc, m_head, idx))) {
1701 			if (error == EFBIG) {
1702 				printf("%s: Tx packet consumes too many "
1703 				    "DMA segments, dropping...\n",
1704 				    device_xname(sc->sc_dev));
1705 				IFQ_DEQUEUE(&ifp->if_snd, m_head);
1706 				m_freem(m_head);
1707 				continue;
1708 			}
1709 
1710 			/*
1711 			 * Short on resources, just stop for now.
1712 			 */
1713 			if (error == ENOBUFS)
1714 				ifp->if_flags |= IFF_OACTIVE;
1715 			break;
1716 		}
1717 
1718 		IFQ_DEQUEUE(&ifp->if_snd, m_head);
1719 
1720 		/*
1721 		 * WE ARE NOW COMMITTED TO TRANSMITTING THE PACKET.
1722 		 */
1723 
1724 		sc->sc_txdescs[pidx].td_frag[0].tf_buflen |=
1725 		    htole16(VGE_TXDESC_Q);
1726 		VGE_TXFRAGSYNC(sc, pidx, 1,
1727 		    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1728 
1729 		if (txs->txs_mbuf != m_head) {
1730 			m_freem(m_head);
1731 			m_head = txs->txs_mbuf;
1732 		}
1733 
1734 		pidx = idx;
1735 		idx = VGE_NEXT_TXDESC(idx);
1736 
1737 		/*
1738 		 * If there's a BPF listener, bounce a copy of this frame
1739 		 * to him.
1740 		 */
1741 		bpf_mtap(ifp, m_head);
1742 	}
1743 
1744 	if (sc->sc_tx_free < ofree) {
1745 		/* TX packet queued */
1746 
1747 		sc->sc_tx_prodidx = idx;
1748 
1749 		/* Issue a transmit command. */
1750 		CSR_WRITE_2(sc, VGE_TXQCSRS, VGE_TXQCSR_WAK0);
1751 
1752 		/*
1753 		 * Use the countdown timer for interrupt moderation.
1754 		 * 'TX done' interrupts are disabled. Instead, we reset the
1755 		 * countdown timer, which will begin counting until it hits
1756 		 * the value in the SSTIMER register, and then trigger an
1757 		 * interrupt. Each time we set the TIMER0_ENABLE bit, the
1758 		 * the timer count is reloaded. Only when the transmitter
1759 		 * is idle will the timer hit 0 and an interrupt fire.
1760 		 */
1761 		CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_TIMER0_ENABLE);
1762 
1763 		/*
1764 		 * Set a timeout in case the chip goes out to lunch.
1765 		 */
1766 		ifp->if_timer = 5;
1767 	}
1768 }
1769 
1770 static int
1771 vge_init(struct ifnet *ifp)
1772 {
1773 	struct vge_softc *sc;
1774 	int i, rc = 0;
1775 
1776 	sc = ifp->if_softc;
1777 
1778 	/*
1779 	 * Cancel pending I/O and free all RX/TX buffers.
1780 	 */
1781 	vge_stop(ifp, 0);
1782 	vge_reset(sc);
1783 
1784 	/* Initialize the RX descriptors and mbufs. */
1785 	memset(sc->sc_rxdescs, 0, sizeof(sc->sc_rxdescs));
1786 	sc->sc_rx_consumed = 0;
1787 	for (i = 0; i < VGE_NRXDESC; i++) {
1788 		if (vge_newbuf(sc, i, NULL) == ENOBUFS) {
1789 			printf("%s: unable to allocate or map rx buffer\n",
1790 			    device_xname(sc->sc_dev));
1791 			return 1; /* XXX */
1792 		}
1793 	}
1794 	sc->sc_rx_prodidx = 0;
1795 	sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
1796 
1797 	/* Initialize the  TX descriptors and mbufs. */
1798 	memset(sc->sc_txdescs, 0, sizeof(sc->sc_txdescs));
1799 	bus_dmamap_sync(sc->sc_dmat, sc->sc_cddmamap,
1800 	    VGE_CDTXOFF(0), sizeof(sc->sc_txdescs),
1801 	    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
1802 	for (i = 0; i < VGE_NTXDESC; i++)
1803 		sc->sc_txsoft[i].txs_mbuf = NULL;
1804 
1805 	sc->sc_tx_prodidx = 0;
1806 	sc->sc_tx_considx = 0;
1807 	sc->sc_tx_free = VGE_NTXDESC;
1808 
1809 	/* Set our station address */
1810 	for (i = 0; i < ETHER_ADDR_LEN; i++)
1811 		CSR_WRITE_1(sc, VGE_PAR0 + i, sc->sc_eaddr[i]);
1812 
1813 	/*
1814 	 * Set receive FIFO threshold. Also allow transmission and
1815 	 * reception of VLAN tagged frames.
1816 	 */
1817 	CSR_CLRBIT_1(sc, VGE_RXCFG, VGE_RXCFG_FIFO_THR|VGE_RXCFG_VTAGOPT);
1818 	CSR_SETBIT_1(sc, VGE_RXCFG, VGE_RXFIFOTHR_128BYTES|VGE_VTAG_OPT2);
1819 
1820 	/* Set DMA burst length */
1821 	CSR_CLRBIT_1(sc, VGE_DMACFG0, VGE_DMACFG0_BURSTLEN);
1822 	CSR_SETBIT_1(sc, VGE_DMACFG0, VGE_DMABURST_128);
1823 
1824 	CSR_SETBIT_1(sc, VGE_TXCFG, VGE_TXCFG_ARB_PRIO|VGE_TXCFG_NONBLK);
1825 
1826 	/* Set collision backoff algorithm */
1827 	CSR_CLRBIT_1(sc, VGE_CHIPCFG1, VGE_CHIPCFG1_CRANDOM|
1828 	    VGE_CHIPCFG1_CAP|VGE_CHIPCFG1_MBA|VGE_CHIPCFG1_BAKOPT);
1829 	CSR_SETBIT_1(sc, VGE_CHIPCFG1, VGE_CHIPCFG1_OFSET);
1830 
1831 	/* Disable LPSEL field in priority resolution */
1832 	CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_LPSEL_DIS);
1833 
1834 	/*
1835 	 * Load the addresses of the DMA queues into the chip.
1836 	 * Note that we only use one transmit queue.
1837 	 */
1838 
1839 	CSR_WRITE_4(sc, VGE_TXDESC_ADDR_LO0, VGE_ADDR_LO(VGE_CDTXADDR(sc, 0)));
1840 	CSR_WRITE_2(sc, VGE_TXDESCNUM, VGE_NTXDESC - 1);
1841 
1842 	CSR_WRITE_4(sc, VGE_RXDESC_ADDR_LO, VGE_ADDR_LO(VGE_CDRXADDR(sc, 0)));
1843 	CSR_WRITE_2(sc, VGE_RXDESCNUM, VGE_NRXDESC - 1);
1844 	CSR_WRITE_2(sc, VGE_RXDESC_RESIDUECNT, VGE_NRXDESC);
1845 
1846 	/* Enable and wake up the RX descriptor queue */
1847 	CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_RUN);
1848 	CSR_WRITE_1(sc, VGE_RXQCSRS, VGE_RXQCSR_WAK);
1849 
1850 	/* Enable the TX descriptor queue */
1851 	CSR_WRITE_2(sc, VGE_TXQCSRS, VGE_TXQCSR_RUN0);
1852 
1853 	/* Set up the receive filter -- allow large frames for VLANs. */
1854 	CSR_WRITE_1(sc, VGE_RXCTL, VGE_RXCTL_RX_UCAST|VGE_RXCTL_RX_GIANT);
1855 
1856 	/* If we want promiscuous mode, set the allframes bit. */
1857 	if (ifp->if_flags & IFF_PROMISC) {
1858 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_PROMISC);
1859 	}
1860 
1861 	/* Set capture broadcast bit to capture broadcast frames. */
1862 	if (ifp->if_flags & IFF_BROADCAST) {
1863 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_BCAST);
1864 	}
1865 
1866 	/* Set multicast bit to capture multicast frames. */
1867 	if (ifp->if_flags & IFF_MULTICAST) {
1868 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_MCAST);
1869 	}
1870 
1871 	/* Init the cam filter. */
1872 	vge_cam_clear(sc);
1873 
1874 	/* Init the multicast filter. */
1875 	vge_setmulti(sc);
1876 
1877 	/* Enable flow control */
1878 
1879 	CSR_WRITE_1(sc, VGE_CRS2, 0x8B);
1880 
1881 	/* Enable jumbo frame reception (if desired) */
1882 
1883 	/* Start the MAC. */
1884 	CSR_WRITE_1(sc, VGE_CRC0, VGE_CR0_STOP);
1885 	CSR_WRITE_1(sc, VGE_CRS1, VGE_CR1_NOPOLL);
1886 	CSR_WRITE_1(sc, VGE_CRS0,
1887 	    VGE_CR0_TX_ENABLE|VGE_CR0_RX_ENABLE|VGE_CR0_START);
1888 
1889 	/*
1890 	 * Configure one-shot timer for microsecond
1891 	 * resulution and load it for 500 usecs.
1892 	 */
1893 	CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_TIMER0_RES);
1894 	CSR_WRITE_2(sc, VGE_SSTIMER, 400);
1895 
1896 	/*
1897 	 * Configure interrupt moderation for receive. Enable
1898 	 * the holdoff counter and load it, and set the RX
1899 	 * suppression count to the number of descriptors we
1900 	 * want to allow before triggering an interrupt.
1901 	 * The holdoff timer is in units of 20 usecs.
1902 	 */
1903 
1904 #ifdef notyet
1905 	CSR_WRITE_1(sc, VGE_INTCTL1, VGE_INTCTL_TXINTSUP_DISABLE);
1906 	/* Select the interrupt holdoff timer page. */
1907 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1908 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_INTHLDOFF);
1909 	CSR_WRITE_1(sc, VGE_INTHOLDOFF, 10); /* ~200 usecs */
1910 
1911 	/* Enable use of the holdoff timer. */
1912 	CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_HOLDOFF);
1913 	CSR_WRITE_1(sc, VGE_INTCTL1, VGE_INTCTL_SC_RELOAD);
1914 
1915 	/* Select the RX suppression threshold page. */
1916 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1917 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_RXSUPPTHR);
1918 	CSR_WRITE_1(sc, VGE_RXSUPPTHR, 64); /* interrupt after 64 packets */
1919 
1920 	/* Restore the page select bits. */
1921 	CSR_CLRBIT_1(sc, VGE_CAMCTL, VGE_CAMCTL_PAGESEL);
1922 	CSR_SETBIT_1(sc, VGE_CAMCTL, VGE_PAGESEL_MAR);
1923 #endif
1924 
1925 #ifdef DEVICE_POLLING
1926 	/*
1927 	 * Disable interrupts if we are polling.
1928 	 */
1929 	if (ifp->if_flags & IFF_POLLING) {
1930 		CSR_WRITE_4(sc, VGE_IMR, 0);
1931 		CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
1932 	} else	/* otherwise ... */
1933 #endif /* DEVICE_POLLING */
1934 	{
1935 	/*
1936 	 * Enable interrupts.
1937 	 */
1938 		CSR_WRITE_4(sc, VGE_IMR, VGE_INTRS);
1939 		CSR_WRITE_4(sc, VGE_ISR, 0);
1940 		CSR_WRITE_1(sc, VGE_CRS3, VGE_CR3_INT_GMSK);
1941 	}
1942 
1943 	if ((rc = ether_mediachange(ifp)) != 0)
1944 		goto out;
1945 
1946 	ifp->if_flags |= IFF_RUNNING;
1947 	ifp->if_flags &= ~IFF_OACTIVE;
1948 
1949 	sc->sc_if_flags = 0;
1950 	sc->sc_link = 0;
1951 
1952 	callout_schedule(&sc->sc_timeout, hz);
1953 
1954 out:
1955 	return rc;
1956 }
1957 
1958 static void
1959 vge_miibus_statchg(struct ifnet *ifp)
1960 {
1961 	struct vge_softc *sc = ifp->if_softc;
1962 	struct mii_data *mii = &sc->sc_mii;
1963 	struct ifmedia_entry *ife = mii->mii_media.ifm_cur;
1964 
1965 	/*
1966 	 * If the user manually selects a media mode, we need to turn
1967 	 * on the forced MAC mode bit in the DIAGCTL register. If the
1968 	 * user happens to choose a full duplex mode, we also need to
1969 	 * set the 'force full duplex' bit. This applies only to
1970 	 * 10Mbps and 100Mbps speeds. In autoselect mode, forced MAC
1971 	 * mode is disabled, and in 1000baseT mode, full duplex is
1972 	 * always implied, so we turn on the forced mode bit but leave
1973 	 * the FDX bit cleared.
1974 	 */
1975 
1976 	switch (IFM_SUBTYPE(ife->ifm_media)) {
1977 	case IFM_AUTO:
1978 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
1979 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1980 		break;
1981 	case IFM_1000_T:
1982 		CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
1983 		CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1984 		break;
1985 	case IFM_100_TX:
1986 	case IFM_10_T:
1987 		CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_MACFORCE);
1988 		if ((ife->ifm_media & IFM_GMASK) == IFM_FDX) {
1989 			CSR_SETBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1990 		} else {
1991 			CSR_CLRBIT_1(sc, VGE_DIAGCTL, VGE_DIAGCTL_FDXFORCE);
1992 		}
1993 		break;
1994 	default:
1995 		printf("%s: unknown media type: %x\n",
1996 		    device_xname(sc->sc_dev),
1997 		    IFM_SUBTYPE(ife->ifm_media));
1998 		break;
1999 	}
2000 }
2001 
2002 static int
2003 vge_ifflags_cb(struct ethercom *ec)
2004 {
2005 	struct ifnet *ifp = &ec->ec_if;
2006 	struct vge_softc *sc = ifp->if_softc;
2007 	int change = ifp->if_flags ^ sc->sc_if_flags;
2008 
2009 	if ((change & ~(IFF_CANTCHANGE|IFF_DEBUG)) != 0)
2010 		return ENETRESET;
2011 	else if ((change & IFF_PROMISC) == 0)
2012 		return 0;
2013 
2014 	if ((ifp->if_flags & IFF_PROMISC) == 0)
2015 		CSR_CLRBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_PROMISC);
2016 	else
2017 		CSR_SETBIT_1(sc, VGE_RXCTL, VGE_RXCTL_RX_PROMISC);
2018 	vge_setmulti(sc);
2019 	return 0;
2020 }
2021 
2022 static int
2023 vge_ioctl(struct ifnet *ifp, u_long command, void *data)
2024 {
2025 	struct vge_softc *sc;
2026 	int s, error;
2027 
2028 	sc = ifp->if_softc;
2029 	error = 0;
2030 
2031 	s = splnet();
2032 
2033 	if ((error = ether_ioctl(ifp, command, data)) == ENETRESET) {
2034 		error = 0;
2035 		if (command != SIOCADDMULTI && command != SIOCDELMULTI)
2036 			;
2037 		else if (ifp->if_flags & IFF_RUNNING) {
2038 			/*
2039 			 * Multicast list has changed; set the hardware filter
2040 			 * accordingly.
2041 			 */
2042 			vge_setmulti(sc);
2043 		}
2044 	}
2045 	sc->sc_if_flags = ifp->if_flags;
2046 
2047 	splx(s);
2048 	return error;
2049 }
2050 
2051 static void
2052 vge_watchdog(struct ifnet *ifp)
2053 {
2054 	struct vge_softc *sc;
2055 	int s;
2056 
2057 	sc = ifp->if_softc;
2058 	s = splnet();
2059 	printf("%s: watchdog timeout\n", device_xname(sc->sc_dev));
2060 	ifp->if_oerrors++;
2061 
2062 	vge_txeof(sc);
2063 	vge_rxeof(sc);
2064 
2065 	vge_init(ifp);
2066 
2067 	splx(s);
2068 }
2069 
2070 /*
2071  * Stop the adapter and free any mbufs allocated to the
2072  * RX and TX lists.
2073  */
2074 static void
2075 vge_stop(struct ifnet *ifp, int disable)
2076 {
2077 	struct vge_softc *sc = ifp->if_softc;
2078 	struct vge_txsoft *txs;
2079 	struct vge_rxsoft *rxs;
2080 	int i, s;
2081 
2082 	s = splnet();
2083 	ifp->if_timer = 0;
2084 
2085 	ifp->if_flags &= ~(IFF_RUNNING | IFF_OACTIVE);
2086 #ifdef DEVICE_POLLING
2087 	ether_poll_deregister(ifp);
2088 #endif /* DEVICE_POLLING */
2089 
2090 	CSR_WRITE_1(sc, VGE_CRC3, VGE_CR3_INT_GMSK);
2091 	CSR_WRITE_1(sc, VGE_CRS0, VGE_CR0_STOP);
2092 	CSR_WRITE_4(sc, VGE_ISR, 0xFFFFFFFF);
2093 	CSR_WRITE_2(sc, VGE_TXQCSRC, 0xFFFF);
2094 	CSR_WRITE_1(sc, VGE_RXQCSRC, 0xFF);
2095 	CSR_WRITE_4(sc, VGE_RXDESC_ADDR_LO, 0);
2096 
2097 	if (sc->sc_rx_mhead != NULL) {
2098 		m_freem(sc->sc_rx_mhead);
2099 		sc->sc_rx_mhead = sc->sc_rx_mtail = NULL;
2100 	}
2101 
2102 	/* Free the TX list buffers. */
2103 
2104 	for (i = 0; i < VGE_NTXDESC; i++) {
2105 		txs = &sc->sc_txsoft[i];
2106 		if (txs->txs_mbuf != NULL) {
2107 			bus_dmamap_unload(sc->sc_dmat, txs->txs_dmamap);
2108 			m_freem(txs->txs_mbuf);
2109 			txs->txs_mbuf = NULL;
2110 		}
2111 	}
2112 
2113 	/* Free the RX list buffers. */
2114 
2115 	for (i = 0; i < VGE_NRXDESC; i++) {
2116 		rxs = &sc->sc_rxsoft[i];
2117 		if (rxs->rxs_mbuf != NULL) {
2118 			bus_dmamap_unload(sc->sc_dmat, rxs->rxs_dmamap);
2119 			m_freem(rxs->rxs_mbuf);
2120 			rxs->rxs_mbuf = NULL;
2121 		}
2122 	}
2123 
2124 	splx(s);
2125 }
2126 
2127 #if VGE_POWER_MANAGEMENT
2128 /*
2129  * Device suspend routine.  Stop the interface and save some PCI
2130  * settings in case the BIOS doesn't restore them properly on
2131  * resume.
2132  */
2133 static int
2134 vge_suspend(device_t dev)
2135 {
2136 	struct vge_softc *sc;
2137 	int i;
2138 
2139 	sc = device_get_softc(dev);
2140 
2141 	vge_stop(sc);
2142 
2143         for (i = 0; i < 5; i++)
2144 		sc->sc_saved_maps[i] =
2145 		    pci_read_config(dev, PCIR_MAPS + i * 4, 4);
2146 	sc->sc_saved_biosaddr = pci_read_config(dev, PCIR_BIOS, 4);
2147 	sc->sc_saved_intline = pci_read_config(dev, PCIR_INTLINE, 1);
2148 	sc->sc_saved_cachelnsz = pci_read_config(dev, PCIR_CACHELNSZ, 1);
2149 	sc->sc_saved_lattimer = pci_read_config(dev, PCIR_LATTIMER, 1);
2150 
2151 	sc->suspended = 1;
2152 
2153 	return 0;
2154 }
2155 
2156 /*
2157  * Device resume routine.  Restore some PCI settings in case the BIOS
2158  * doesn't, re-enable busmastering, and restart the interface if
2159  * appropriate.
2160  */
2161 static int
2162 vge_resume(device_t dev)
2163 {
2164 	struct vge_softc *sc;
2165 	struct ifnet *ifp;
2166 	int i;
2167 
2168 	sc = device_private(dev);
2169 	ifp = &sc->sc_ethercom.ec_if;
2170 
2171         /* better way to do this? */
2172 	for (i = 0; i < 5; i++)
2173 		pci_write_config(dev, PCIR_MAPS + i * 4,
2174 		    sc->sc_saved_maps[i], 4);
2175 	pci_write_config(dev, PCIR_BIOS, sc->sc_saved_biosaddr, 4);
2176 	pci_write_config(dev, PCIR_INTLINE, sc->sc_saved_intline, 1);
2177 	pci_write_config(dev, PCIR_CACHELNSZ, sc->sc_saved_cachelnsz, 1);
2178 	pci_write_config(dev, PCIR_LATTIMER, sc->sc_saved_lattimer, 1);
2179 
2180 	/* reenable busmastering */
2181 	pci_enable_busmaster(dev);
2182 	pci_enable_io(dev, SYS_RES_MEMORY);
2183 
2184 	/* reinitialize interface if necessary */
2185 	if (ifp->if_flags & IFF_UP)
2186 		vge_init(sc);
2187 
2188 	sc->suspended = 0;
2189 
2190 	return 0;
2191 }
2192 #endif
2193 
2194 /*
2195  * Stop all chip I/O so that the kernel's probe routines don't
2196  * get confused by errant DMAs when rebooting.
2197  */
2198 static bool
2199 vge_shutdown(device_t self, int howto)
2200 {
2201 	struct vge_softc *sc;
2202 
2203 	sc = device_private(self);
2204 	vge_stop(&sc->sc_ethercom.ec_if, 1);
2205 
2206 	return true;
2207 }
2208