xref: /netbsd-src/sys/arch/arm/broadcom/bcm53xx_eth.c (revision d16b7486a53dcb8072b60ec6fcb4373a2d0c27b7)
1 /*-
2  * Copyright (c) 2012 The NetBSD Foundation, Inc.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to The NetBSD Foundation
6  * by Matt Thomas of 3am Software Foundry.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
18  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
19  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
21  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
22  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
23  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
24  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
25  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
26  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27  * POSSIBILITY OF SUCH DAMAGE.
28  */
29 
30 #define _ARM32_BUS_DMA_PRIVATE
31 #define GMAC_PRIVATE
32 
33 #include "locators.h"
34 #include "opt_broadcom.h"
35 
36 #include <sys/cdefs.h>
37 
38 __KERNEL_RCSID(1, "$NetBSD: bcm53xx_eth.c,v 1.42 2022/09/17 19:41:18 thorpej Exp $");
39 
40 #include <sys/param.h>
41 #include <sys/atomic.h>
42 #include <sys/bus.h>
43 #include <sys/device.h>
44 #include <sys/ioctl.h>
45 #include <sys/intr.h>
46 #include <sys/kmem.h>
47 #include <sys/mutex.h>
48 #include <sys/socket.h>
49 #include <sys/systm.h>
50 #include <sys/workqueue.h>
51 
52 #include <net/if.h>
53 #include <net/if_ether.h>
54 #include <net/if_media.h>
55 #include <net/if_dl.h>
56 #include <net/bpf.h>
57 
58 #include <dev/mii/miivar.h>
59 
60 #include <arm/locore.h>
61 
62 #include <arm/broadcom/bcm53xx_reg.h>
63 #include <arm/broadcom/bcm53xx_var.h>
64 
65 //#define BCMETH_MPSAFE
66 
67 #ifdef BCMETH_COUNTERS
68 #define	BCMETH_EVCNT_ADD(a, b)	((void)((a).ev_count += (b)))
69 #else
70 #define	BCMETH_EVCNT_ADD(a, b)	do { } while (/*CONSTCOND*/0)
71 #endif
72 #define	BCMETH_EVCNT_INCR(a)	BCMETH_EVCNT_ADD((a), 1)
73 
74 #define	BCMETH_MAXTXMBUFS	128
75 #define	BCMETH_NTXSEGS		30
76 #define	BCMETH_MAXRXMBUFS	255
77 #define	BCMETH_MINRXMBUFS	64
78 #define	BCMETH_NRXSEGS		1
79 #define	BCMETH_RINGSIZE		PAGE_SIZE
80 
81 #if 1
82 #define	BCMETH_RCVMAGIC		0xfeedface
83 #endif
84 
85 static int bcmeth_ccb_match(device_t, cfdata_t, void *);
86 static void bcmeth_ccb_attach(device_t, device_t, void *);
87 
88 struct bcmeth_txqueue {
89 	bus_dmamap_t txq_descmap;
90 	struct gmac_txdb *txq_consumer;
91 	struct gmac_txdb *txq_producer;
92 	struct gmac_txdb *txq_first;
93 	struct gmac_txdb *txq_last;
94 	struct ifqueue txq_mbufs;
95 	struct mbuf *txq_next;
96 	size_t txq_free;
97 	size_t txq_threshold;
98 	size_t txq_lastintr;
99 	bus_size_t txq_reg_xmtaddrlo;
100 	bus_size_t txq_reg_xmtptr;
101 	bus_size_t txq_reg_xmtctl;
102 	bus_size_t txq_reg_xmtsts0;
103 	bus_size_t txq_reg_xmtsts1;
104 	bus_dma_segment_t txq_descmap_seg;
105 };
106 
107 struct bcmeth_rxqueue {
108 	bus_dmamap_t rxq_descmap;
109 	struct gmac_rxdb *rxq_consumer;
110 	struct gmac_rxdb *rxq_producer;
111 	struct gmac_rxdb *rxq_first;
112 	struct gmac_rxdb *rxq_last;
113 	struct mbuf *rxq_mhead;
114 	struct mbuf **rxq_mtail;
115 	struct mbuf *rxq_mconsumer;
116 	size_t rxq_inuse;
117 	size_t rxq_threshold;
118 	bus_size_t rxq_reg_rcvaddrlo;
119 	bus_size_t rxq_reg_rcvptr;
120 	bus_size_t rxq_reg_rcvctl;
121 	bus_size_t rxq_reg_rcvsts0;
122 	bus_size_t rxq_reg_rcvsts1;
123 	bus_dma_segment_t rxq_descmap_seg;
124 };
125 
126 struct bcmeth_mapcache {
127 	u_int dmc_nmaps;
128 	u_int dmc_maxseg;
129 	u_int dmc_maxmaps;
130 	u_int dmc_maxmapsize;
131 	bus_dmamap_t dmc_maps[0];
132 };
133 
134 struct bcmeth_softc {
135 	device_t sc_dev;
136 	bus_space_tag_t sc_bst;
137 	bus_space_handle_t sc_bsh;
138 	bus_dma_tag_t sc_dmat;
139 	kmutex_t *sc_lock;
140 	kmutex_t *sc_hwlock;
141 	struct ethercom sc_ec;
142 #define	sc_if		sc_ec.ec_if
143 	struct ifmedia sc_media;
144 	void *sc_soft_ih;
145 	void *sc_ih;
146 
147 	struct bcmeth_rxqueue sc_rxq;
148 	struct bcmeth_txqueue sc_txq;
149 
150 	size_t sc_rcvoffset;
151 	uint32_t sc_macaddr[2];
152 	uint32_t sc_maxfrm;
153 	uint32_t sc_cmdcfg;
154 	uint32_t sc_intmask;
155 	uint32_t sc_rcvlazy;
156 	volatile uint32_t sc_soft_flags;
157 #define	SOFT_RXINTR		0x01
158 #define	SOFT_TXINTR		0x02
159 
160 #ifdef BCMETH_COUNTERS
161 	struct evcnt sc_ev_intr;
162 	struct evcnt sc_ev_soft_intr;
163 	struct evcnt sc_ev_work;
164 	struct evcnt sc_ev_tx_stall;
165 	struct evcnt sc_ev_rx_badmagic_lo;
166 	struct evcnt sc_ev_rx_badmagic_hi;
167 #endif
168 
169 	struct ifqueue sc_rx_bufcache;
170 	struct bcmeth_mapcache *sc_rx_mapcache;
171 	struct bcmeth_mapcache *sc_tx_mapcache;
172 
173 	struct workqueue *sc_workq;
174 	struct work sc_work;
175 
176 	volatile uint32_t sc_work_flags;
177 #define	WORK_RXINTR		0x01
178 #define	WORK_RXUNDERFLOW	0x02
179 #define	WORK_REINIT		0x04
180 
181 	uint8_t sc_enaddr[ETHER_ADDR_LEN];
182 };
183 
184 static void bcmeth_ifstart(struct ifnet *);
185 static void bcmeth_ifwatchdog(struct ifnet *);
186 static int bcmeth_ifinit(struct ifnet *);
187 static void bcmeth_ifstop(struct ifnet *, int);
188 static int bcmeth_ifioctl(struct ifnet *, u_long, void *);
189 
190 static int bcmeth_mapcache_create(struct bcmeth_softc *,
191     struct bcmeth_mapcache **, size_t, size_t, size_t);
192 static void bcmeth_mapcache_destroy(struct bcmeth_softc *,
193     struct bcmeth_mapcache *);
194 static bus_dmamap_t bcmeth_mapcache_get(struct bcmeth_softc *,
195     struct bcmeth_mapcache *);
196 static void bcmeth_mapcache_put(struct bcmeth_softc *,
197     struct bcmeth_mapcache *, bus_dmamap_t);
198 
199 static int bcmeth_txq_attach(struct bcmeth_softc *,
200     struct bcmeth_txqueue *, u_int);
201 static void bcmeth_txq_purge(struct bcmeth_softc *,
202     struct bcmeth_txqueue *);
203 static void bcmeth_txq_reset(struct bcmeth_softc *,
204     struct bcmeth_txqueue *);
205 static bool bcmeth_txq_consume(struct bcmeth_softc *,
206     struct bcmeth_txqueue *);
207 static bool bcmeth_txq_produce(struct bcmeth_softc *,
208     struct bcmeth_txqueue *, struct mbuf *m);
209 static bool bcmeth_txq_active_p(struct bcmeth_softc *,
210     struct bcmeth_txqueue *);
211 
212 static int bcmeth_rxq_attach(struct bcmeth_softc *,
213     struct bcmeth_rxqueue *, u_int);
214 static bool bcmeth_rxq_produce(struct bcmeth_softc *,
215     struct bcmeth_rxqueue *);
216 static void bcmeth_rxq_purge(struct bcmeth_softc *,
217     struct bcmeth_rxqueue *, bool);
218 static void bcmeth_rxq_reset(struct bcmeth_softc *,
219     struct bcmeth_rxqueue *);
220 
221 static int bcmeth_intr(void *);
222 #ifdef BCMETH_MPSAFETX
223 static void bcmeth_soft_txintr(struct bcmeth_softc *);
224 #endif
225 static void bcmeth_soft_intr(void *);
226 static void bcmeth_worker(struct work *, void *);
227 
228 static int bcmeth_mediachange(struct ifnet *);
229 static void bcmeth_mediastatus(struct ifnet *, struct ifmediareq *);
230 
231 static inline uint32_t
232 bcmeth_read_4(struct bcmeth_softc *sc, bus_size_t o)
233 {
234 	return bus_space_read_4(sc->sc_bst, sc->sc_bsh, o);
235 }
236 
237 static inline void
238 bcmeth_write_4(struct bcmeth_softc *sc, bus_size_t o, uint32_t v)
239 {
240 	bus_space_write_4(sc->sc_bst, sc->sc_bsh, o, v);
241 }
242 
243 CFATTACH_DECL_NEW(bcmeth_ccb, sizeof(struct bcmeth_softc),
244 	bcmeth_ccb_match, bcmeth_ccb_attach, NULL, NULL);
245 
246 static int
247 bcmeth_ccb_match(device_t parent, cfdata_t cf, void *aux)
248 {
249 	struct bcmccb_attach_args * const ccbaa = aux;
250 	const struct bcm_locators * const loc = &ccbaa->ccbaa_loc;
251 
252 	if (strcmp(cf->cf_name, loc->loc_name))
253 		return 0;
254 
255 #ifdef DIAGNOSTIC
256 	const int port = cf->cf_loc[BCMCCBCF_PORT];
257 #endif
258 	KASSERT(port == BCMCCBCF_PORT_DEFAULT || port == loc->loc_port);
259 
260 	return 1;
261 }
262 
263 static void
264 bcmeth_ccb_attach(device_t parent, device_t self, void *aux)
265 {
266 	struct bcmeth_softc * const sc = device_private(self);
267 	struct ethercom * const ec = &sc->sc_ec;
268 	struct ifnet * const ifp = &ec->ec_if;
269 	struct bcmccb_attach_args * const ccbaa = aux;
270 	const struct bcm_locators * const loc = &ccbaa->ccbaa_loc;
271 	const char * const xname = device_xname(self);
272 	prop_dictionary_t dict = device_properties(self);
273 	int error;
274 
275 	sc->sc_bst = ccbaa->ccbaa_ccb_bst;
276 	sc->sc_dmat = ccbaa->ccbaa_dmat;
277 	bus_space_subregion(sc->sc_bst, ccbaa->ccbaa_ccb_bsh,
278 	    loc->loc_offset, loc->loc_size, &sc->sc_bsh);
279 
280 	/*
281 	 * We need to use the coherent dma tag for the GMAC.
282 	 */
283 	sc->sc_dmat = &bcm53xx_coherent_dma_tag;
284 #if _ARM32_NEED_BUS_DMA_BOUNCE
285 	if (device_cfdata(self)->cf_flags & 2) {
286 		sc->sc_dmat = &bcm53xx_bounce_dma_tag;
287 	}
288 #endif
289 
290 	prop_data_t eaprop = prop_dictionary_get(dict, "mac-address");
291 	if (eaprop == NULL) {
292 		uint32_t mac0 = bcmeth_read_4(sc, UNIMAC_MAC_0);
293 		uint32_t mac1 = bcmeth_read_4(sc, UNIMAC_MAC_1);
294 		if ((mac0 == 0 && mac1 == 0) || (mac1 & 1)) {
295 			aprint_error(": mac-address property is missing\n");
296 			return;
297 		}
298 		sc->sc_enaddr[0] = (mac0 >> 0) & 0xff;
299 		sc->sc_enaddr[1] = (mac0 >> 8) & 0xff;
300 		sc->sc_enaddr[2] = (mac0 >> 16) & 0xff;
301 		sc->sc_enaddr[3] = (mac0 >> 24) & 0xff;
302 		sc->sc_enaddr[4] = (mac1 >> 0) & 0xff;
303 		sc->sc_enaddr[5] = (mac1 >> 8) & 0xff;
304 	} else {
305 		KASSERT(prop_object_type(eaprop) == PROP_TYPE_DATA);
306 		KASSERT(prop_data_size(eaprop) == ETHER_ADDR_LEN);
307 		memcpy(sc->sc_enaddr, prop_data_data_nocopy(eaprop),
308 		    ETHER_ADDR_LEN);
309 	}
310 	sc->sc_dev = self;
311 	sc->sc_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SOFTNET);
312 	sc->sc_hwlock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_VM);
313 
314 	bcmeth_write_4(sc, GMAC_INTMASK, 0);	// disable interrupts
315 
316 	aprint_naive("\n");
317 	aprint_normal(": Gigabit Ethernet Controller\n");
318 
319 	error = bcmeth_rxq_attach(sc, &sc->sc_rxq, 0);
320 	if (error) {
321 		aprint_error(": failed to init rxq: %d\n", error);
322 		goto fail_1;
323 	}
324 
325 	error = bcmeth_txq_attach(sc, &sc->sc_txq, 0);
326 	if (error) {
327 		aprint_error(": failed to init txq: %d\n", error);
328 		goto fail_1;
329 	}
330 
331 	error = bcmeth_mapcache_create(sc, &sc->sc_rx_mapcache,
332 	    BCMETH_MAXRXMBUFS, MCLBYTES, BCMETH_NRXSEGS);
333 	if (error) {
334 		aprint_error(": failed to allocate rx dmamaps: %d\n", error);
335 		goto fail_1;
336 	}
337 
338 	error = bcmeth_mapcache_create(sc, &sc->sc_tx_mapcache,
339 	    BCMETH_MAXTXMBUFS, MCLBYTES, BCMETH_NTXSEGS);
340 	if (error) {
341 		aprint_error(": failed to allocate tx dmamaps: %d\n", error);
342 		goto fail_1;
343 	}
344 
345 	error = workqueue_create(&sc->sc_workq, xname, bcmeth_worker, sc,
346 	    (PRI_USER + MAXPRI_USER) / 2, IPL_NET, WQ_MPSAFE|WQ_PERCPU);
347 	if (error) {
348 		aprint_error(": failed to create workqueue: %d\n", error);
349 		goto fail_2;
350 	}
351 
352 	sc->sc_soft_ih = softint_establish(SOFTINT_MPSAFE | SOFTINT_NET,
353 	    bcmeth_soft_intr, sc);
354 
355 	if (sc->sc_ih == NULL) {
356 		aprint_error_dev(self, "failed to establish interrupt %d\n",
357 		     loc->loc_intrs[0]);
358 		goto fail_3;
359 	}
360 
361 	sc->sc_ih = intr_establish(loc->loc_intrs[0], IPL_VM, IST_LEVEL,
362 	    bcmeth_intr, sc);
363 
364 	if (sc->sc_ih == NULL) {
365 		aprint_error_dev(self, "failed to establish interrupt %d\n",
366 		     loc->loc_intrs[0]);
367 		goto fail_4;
368 	} else {
369 		aprint_normal_dev(self, "interrupting on irq %d\n",
370 		     loc->loc_intrs[0]);
371 	}
372 
373 	aprint_normal_dev(sc->sc_dev, "Ethernet address %s\n",
374 	    ether_sprintf(sc->sc_enaddr));
375 
376 	/*
377 	 * Since each port in plugged into the switch/flow-accelerator,
378 	 * we hard code at Gige Full-Duplex with Flow Control enabled.
379 	 */
380 	int ifmedia = IFM_ETHER | IFM_1000_T | IFM_FDX;
381 	//ifmedia |= IFM_FLOW | IFM_ETH_TXPAUSE | IFM_ETH_RXPAUSE;
382 	ec->ec_ifmedia = &sc->sc_media;
383 	ifmedia_init(&sc->sc_media, IFM_IMASK, bcmeth_mediachange,
384 	    bcmeth_mediastatus);
385 	ifmedia_add(&sc->sc_media, ifmedia, 0, NULL);
386 	ifmedia_set(&sc->sc_media, ifmedia);
387 
388 	ec->ec_capabilities = ETHERCAP_VLAN_MTU | ETHERCAP_JUMBO_MTU;
389 
390 	strlcpy(ifp->if_xname, xname, IFNAMSIZ);
391 	ifp->if_softc = sc;
392 	ifp->if_baudrate = IF_Mbps(1000);
393 	ifp->if_capabilities = 0;
394 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
395 #ifdef BCMETH_MPSAFE
396 	ifp->if_flags2 = IFF2_MPSAFE;
397 #endif
398 	ifp->if_ioctl = bcmeth_ifioctl;
399 	ifp->if_start = bcmeth_ifstart;
400 	ifp->if_watchdog = bcmeth_ifwatchdog;
401 	ifp->if_init = bcmeth_ifinit;
402 	ifp->if_stop = bcmeth_ifstop;
403 	IFQ_SET_READY(&ifp->if_snd);
404 
405 	bcmeth_ifstop(ifp, true);
406 
407 	/*
408 	 * Attach the interface.
409 	 */
410 	if_initialize(ifp);
411 	ether_ifattach(ifp, sc->sc_enaddr);
412 	if_register(ifp);
413 
414 #ifdef BCMETH_COUNTERS
415 	evcnt_attach_dynamic(&sc->sc_ev_intr, EVCNT_TYPE_INTR,
416 	    NULL, xname, "intr");
417 	evcnt_attach_dynamic(&sc->sc_ev_soft_intr, EVCNT_TYPE_INTR,
418 	    NULL, xname, "soft intr");
419 	evcnt_attach_dynamic(&sc->sc_ev_work, EVCNT_TYPE_MISC,
420 	    NULL, xname, "work items");
421 	evcnt_attach_dynamic(&sc->sc_ev_tx_stall, EVCNT_TYPE_MISC,
422 	    NULL, xname, "tx stalls");
423 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_lo, EVCNT_TYPE_MISC,
424 	    NULL, xname, "rx badmagic lo");
425 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_hi, EVCNT_TYPE_MISC,
426 	    NULL, xname, "rx badmagic hi");
427 #endif
428 
429 	return;
430 
431 fail_4:
432 	intr_disestablish(sc->sc_ih);
433 fail_3:
434 	softint_disestablish(sc->sc_soft_ih);
435 fail_2:
436 	workqueue_destroy(sc->sc_workq);
437 fail_1:
438 	mutex_obj_free(sc->sc_lock);
439 	mutex_obj_free(sc->sc_hwlock);
440 }
441 
442 static int
443 bcmeth_mediachange(struct ifnet *ifp)
444 {
445 	//struct bcmeth_softc * const sc = ifp->if_softc;
446 	return 0;
447 }
448 
449 static void
450 bcmeth_mediastatus(struct ifnet *ifp, struct ifmediareq *ifm)
451 {
452 	//struct bcmeth_softc * const sc = ifp->if_softc;
453 
454 	ifm->ifm_status = IFM_AVALID | IFM_ACTIVE;
455 	ifm->ifm_active = IFM_ETHER | IFM_FDX | IFM_1000_T;
456 }
457 
458 static uint64_t
459 bcmeth_macaddr_create(const uint8_t *enaddr)
460 {
461 	return (enaddr[3] << 0)			// UNIMAC_MAC_0
462 	    |  (enaddr[2] << 8)			// UNIMAC_MAC_0
463 	    |  (enaddr[1] << 16)		// UNIMAC_MAC_0
464 	    |  ((uint64_t)enaddr[0] << 24)	// UNIMAC_MAC_0
465 	    |  ((uint64_t)enaddr[5] << 32)	// UNIMAC_MAC_1
466 	    |  ((uint64_t)enaddr[4] << 40);	// UNIMAC_MAC_1
467 }
468 
469 static int
470 bcmeth_ifinit(struct ifnet *ifp)
471 {
472 	struct bcmeth_softc * const sc = ifp->if_softc;
473 	int error = 0;
474 
475 	sc->sc_maxfrm = uimax(ifp->if_mtu + 32, MCLBYTES);
476 	if (ifp->if_mtu > ETHERMTU_JUMBO)
477 		return error;
478 
479 	KASSERT(ifp->if_flags & IFF_UP);
480 
481 	/*
482 	 * Stop the interface
483 	 */
484 	bcmeth_ifstop(ifp, 0);
485 
486 	/*
487 	 * Reserve enough space at the front so that we can insert a maxsized
488 	 * link header and a VLAN tag.  Also make sure we have enough room for
489 	 * the rcvsts field as well.
490 	 */
491 	KASSERT(ALIGN(max_linkhdr) == max_linkhdr);
492 	KASSERTMSG(max_linkhdr > sizeof(struct ether_header), "%u > %zu",
493 	    max_linkhdr, sizeof(struct ether_header));
494 	sc->sc_rcvoffset = max_linkhdr + 4 - sizeof(struct ether_header);
495 	if (sc->sc_rcvoffset <= 4)
496 		sc->sc_rcvoffset += 4;
497 	KASSERT((sc->sc_rcvoffset & 3) == 2);
498 	KASSERT(sc->sc_rcvoffset <= __SHIFTOUT(RCVCTL_RCVOFFSET, RCVCTL_RCVOFFSET));
499 	KASSERT(sc->sc_rcvoffset >= 6);
500 
501 	/*
502 	 * If our frame size has changed (or it's our first time through)
503 	 * destroy the existing transmit mapcache.
504 	 */
505 	if (sc->sc_tx_mapcache != NULL
506 	    && sc->sc_maxfrm != sc->sc_tx_mapcache->dmc_maxmapsize) {
507 		bcmeth_mapcache_destroy(sc, sc->sc_tx_mapcache);
508 		sc->sc_tx_mapcache = NULL;
509 	}
510 
511 	if (sc->sc_tx_mapcache == NULL) {
512 		error = bcmeth_mapcache_create(sc, &sc->sc_tx_mapcache,
513 		    BCMETH_MAXTXMBUFS, sc->sc_maxfrm, BCMETH_NTXSEGS);
514 		if (error)
515 			return error;
516 	}
517 
518 	sc->sc_cmdcfg = NO_LENGTH_CHECK | PAUSE_IGNORE
519 	    | __SHIFTIN(ETH_SPEED_1000, ETH_SPEED)
520 	    | RX_ENA | TX_ENA;
521 
522 	if (ifp->if_flags & IFF_PROMISC) {
523 		sc->sc_cmdcfg |= PROMISC_EN;
524 	} else {
525 		sc->sc_cmdcfg &= ~PROMISC_EN;
526 	}
527 
528 	const uint8_t * const lladdr = CLLADDR(ifp->if_sadl);
529 	const uint64_t macstnaddr = bcmeth_macaddr_create(lladdr);
530 
531 	/*
532 	 * We make sure that a received Ethernet packet start on a non-word
533 	 * boundary so that the packet payload will be on a word boundary.
534 	 * So to check the destination address we keep around two words to
535 	 * quickly compare with.
536 	 */
537 #if __ARMEL__
538 	sc->sc_macaddr[0] = lladdr[0] | (lladdr[1] << 8);
539 	sc->sc_macaddr[1] = lladdr[2] | (lladdr[3] << 8)
540 	    | (lladdr[4] << 16) | (lladdr[5] << 24);
541 #else
542 	sc->sc_macaddr[0] = lladdr[1] | (lladdr[0] << 8);
543 	sc->sc_macaddr[1] = lladdr[5] | (lladdr[4] << 8)
544 	    | (lladdr[1] << 16) | (lladdr[2] << 24);
545 #endif
546 
547 	sc->sc_intmask = DESCPROTOERR | DATAERR | DESCERR;
548 
549 	/* 5. Load RCVADDR_LO with new pointer */
550 	bcmeth_rxq_reset(sc, &sc->sc_rxq);
551 
552 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
553 	    __SHIFTIN(sc->sc_rcvoffset, RCVCTL_RCVOFFSET)
554 	    | RCVCTL_PARITY_DIS
555 	    | RCVCTL_OFLOW_CONTINUE
556 	    | __SHIFTIN(3, RCVCTL_BURSTLEN));
557 
558 	/* 6. Load XMTADDR_LO with new pointer */
559 	bcmeth_txq_reset(sc, &sc->sc_txq);
560 
561 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl, XMTCTL_DMA_ACT_INDEX
562 	    | XMTCTL_PARITY_DIS
563 	    | __SHIFTIN(3, XMTCTL_BURSTLEN));
564 
565 	/* 7. Setup other UNIMAC registers */
566 	bcmeth_write_4(sc, UNIMAC_FRAME_LEN, sc->sc_maxfrm);
567 	bcmeth_write_4(sc, UNIMAC_MAC_0, (uint32_t)(macstnaddr >>  0));
568 	bcmeth_write_4(sc, UNIMAC_MAC_1, (uint32_t)(macstnaddr >> 32));
569 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, sc->sc_cmdcfg);
570 
571 	uint32_t devctl = bcmeth_read_4(sc, GMAC_DEVCONTROL);
572 	devctl |= RGMII_LINK_STATUS_SEL | NWAY_AUTO_POLL_EN | TXARB_STRICT_MODE;
573 	devctl &= ~FLOW_CTRL_MODE;
574 	devctl &= ~MIB_RD_RESET_EN;
575 	devctl &= ~RXQ_OVERFLOW_CTRL_SEL;
576 	devctl &= ~CPU_FLOW_CTRL_ON;
577 	bcmeth_write_4(sc, GMAC_DEVCONTROL, devctl);
578 
579 	/* Setup lazy receive (at most 1ms). */
580 	const struct cpu_softc * const cpu = curcpu()->ci_softc;
581 	sc->sc_rcvlazy =  __SHIFTIN(4, INTRCVLAZY_FRAMECOUNT)
582 	     | __SHIFTIN(cpu->cpu_clk.clk_apb / 1000, INTRCVLAZY_TIMEOUT);
583 	bcmeth_write_4(sc, GMAC_INTRCVLAZY, sc->sc_rcvlazy);
584 
585 	/* 11. Enable transmit queues in TQUEUE, and ensure that the transmit scheduling mode is correctly set in TCTRL. */
586 	sc->sc_intmask |= XMTINT_0 | XMTUF;
587 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl,
588 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl) | XMTCTL_ENABLE);
589 
590 
591 	/* 12. Enable receive queues in RQUEUE, */
592 	sc->sc_intmask |= RCVINT | RCVDESCUF | RCVFIFOOF;
593 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
594 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl) | RCVCTL_ENABLE);
595 
596 	bcmeth_rxq_produce(sc, &sc->sc_rxq);	/* fill with rx buffers */
597 
598 #if 0
599 	aprint_normal_dev(sc->sc_dev,
600 	    "devctl=%#x ucmdcfg=%#x xmtctl=%#x rcvctl=%#x\n",
601 	    devctl, sc->sc_cmdcfg,
602 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl),
603 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl));
604 #endif
605 
606 	sc->sc_soft_flags = 0;
607 
608 	bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
609 
610 	ifp->if_flags |= IFF_RUNNING;
611 
612 	return error;
613 }
614 
615 static void
616 bcmeth_ifstop(struct ifnet *ifp, int disable)
617 {
618 	struct bcmeth_softc * const sc = ifp->if_softc;
619 	struct bcmeth_txqueue * const txq = &sc->sc_txq;
620 	struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
621 
622 	KASSERT(!cpu_intr_p());
623 
624 	sc->sc_soft_flags = 0;
625 	sc->sc_work_flags = 0;
626 
627 	/* Disable Rx processing */
628 	bcmeth_write_4(sc, rxq->rxq_reg_rcvctl,
629 	    bcmeth_read_4(sc, rxq->rxq_reg_rcvctl) & ~RCVCTL_ENABLE);
630 
631 	/* Disable Tx processing */
632 	bcmeth_write_4(sc, txq->txq_reg_xmtctl,
633 	    bcmeth_read_4(sc, txq->txq_reg_xmtctl) & ~XMTCTL_ENABLE);
634 
635 	/* Disable all interrupts */
636 	bcmeth_write_4(sc, GMAC_INTMASK, 0);
637 
638 	for (;;) {
639 		uint32_t tx0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
640 		uint32_t rx0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
641 		if (__SHIFTOUT(tx0, XMTSTATE) == XMTSTATE_DIS
642 		    && __SHIFTOUT(rx0, RCVSTATE) == RCVSTATE_DIS)
643 			break;
644 		delay(50);
645 	}
646 	/*
647 	 * Now reset the controller.
648 	 *
649 	 * 3. Set SW_RESET bit in UNIMAC_COMMAND_CONFIG register
650 	 * 4. Clear SW_RESET bit in UNIMAC_COMMAND_CONFIG register
651 	 */
652 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, SW_RESET);
653 	bcmeth_write_4(sc, GMAC_INTSTATUS, ~0);
654 	sc->sc_intmask = 0;
655 	ifp->if_flags &= ~IFF_RUNNING;
656 
657 	/*
658 	 * Let's consume any remaining transmitted packets.  And if we are
659 	 * disabling the interface, purge ourselves of any untransmitted
660 	 * packets.  But don't consume any received packets, just drop them.
661 	 * If we aren't disabling the interface, save the mbufs in the
662 	 * receive queue for reuse.
663 	 */
664 	bcmeth_rxq_purge(sc, &sc->sc_rxq, disable);
665 	bcmeth_txq_consume(sc, &sc->sc_txq);
666 	if (disable) {
667 		bcmeth_txq_purge(sc, &sc->sc_txq);
668 		IF_PURGE(&ifp->if_snd);
669 	}
670 
671 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, 0);
672 }
673 
674 static void
675 bcmeth_ifwatchdog(struct ifnet *ifp)
676 {
677 }
678 
679 static int
680 bcmeth_ifioctl(struct ifnet *ifp, u_long cmd, void *data)
681 {
682 	const int s = splnet();
683 	int error;
684 
685 	switch (cmd) {
686 	default:
687 		error = ether_ioctl(ifp, cmd, data);
688 		if (error != ENETRESET)
689 			break;
690 
691 		if (cmd == SIOCADDMULTI || cmd == SIOCDELMULTI) {
692 			error = 0;
693 			break;
694 		}
695 		error = bcmeth_ifinit(ifp);
696 		break;
697 	}
698 
699 	splx(s);
700 	return error;
701 }
702 
703 static void
704 bcmeth_rxq_desc_presync(
705 	struct bcmeth_softc *sc,
706 	struct bcmeth_rxqueue *rxq,
707 	struct gmac_rxdb *rxdb,
708 	size_t count)
709 {
710 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
711 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
712 	    BUS_DMASYNC_PREWRITE);
713 }
714 
715 static void
716 bcmeth_rxq_desc_postsync(
717 	struct bcmeth_softc *sc,
718 	struct bcmeth_rxqueue *rxq,
719 	struct gmac_rxdb *rxdb,
720 	size_t count)
721 {
722 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
723 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
724 	    BUS_DMASYNC_POSTWRITE);
725 }
726 
727 static void
728 bcmeth_txq_desc_presync(
729 	struct bcmeth_softc *sc,
730 	struct bcmeth_txqueue *txq,
731 	struct gmac_txdb *txdb,
732 	size_t count)
733 {
734 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
735 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
736 	    BUS_DMASYNC_PREWRITE);
737 }
738 
739 static void
740 bcmeth_txq_desc_postsync(
741 	struct bcmeth_softc *sc,
742 	struct bcmeth_txqueue *txq,
743 	struct gmac_txdb *txdb,
744 	size_t count)
745 {
746 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
747 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
748 	    BUS_DMASYNC_POSTWRITE);
749 }
750 
751 static bus_dmamap_t
752 bcmeth_mapcache_get(
753 	struct bcmeth_softc *sc,
754 	struct bcmeth_mapcache *dmc)
755 {
756 	KASSERT(dmc->dmc_nmaps > 0);
757 	KASSERT(dmc->dmc_maps[dmc->dmc_nmaps-1] != NULL);
758 	return dmc->dmc_maps[--dmc->dmc_nmaps];
759 }
760 
761 static void
762 bcmeth_mapcache_put(
763 	struct bcmeth_softc *sc,
764 	struct bcmeth_mapcache *dmc,
765 	bus_dmamap_t map)
766 {
767 	KASSERT(map != NULL);
768 	KASSERT(dmc->dmc_nmaps < dmc->dmc_maxmaps);
769 	dmc->dmc_maps[dmc->dmc_nmaps++] = map;
770 }
771 
772 static void
773 bcmeth_mapcache_destroy(
774 	struct bcmeth_softc *sc,
775 	struct bcmeth_mapcache *dmc)
776 {
777 	const size_t dmc_size =
778 	    offsetof(struct bcmeth_mapcache, dmc_maps[dmc->dmc_maxmaps]);
779 
780 	for (u_int i = 0; i < dmc->dmc_maxmaps; i++) {
781 		bus_dmamap_destroy(sc->sc_dmat, dmc->dmc_maps[i]);
782 	}
783 	kmem_intr_free(dmc, dmc_size);
784 }
785 
786 static int
787 bcmeth_mapcache_create(
788 	struct bcmeth_softc *sc,
789 	struct bcmeth_mapcache **dmc_p,
790 	size_t maxmaps,
791 	size_t maxmapsize,
792 	size_t maxseg)
793 {
794 	const size_t dmc_size =
795 	    offsetof(struct bcmeth_mapcache, dmc_maps[maxmaps]);
796 	struct bcmeth_mapcache * const dmc =
797 		kmem_intr_zalloc(dmc_size, KM_NOSLEEP);
798 
799 	dmc->dmc_maxmaps = maxmaps;
800 	dmc->dmc_nmaps = maxmaps;
801 	dmc->dmc_maxmapsize = maxmapsize;
802 	dmc->dmc_maxseg = maxseg;
803 
804 	for (u_int i = 0; i < maxmaps; i++) {
805 		int error = bus_dmamap_create(sc->sc_dmat, dmc->dmc_maxmapsize,
806 		     dmc->dmc_maxseg, dmc->dmc_maxmapsize, 0,
807 		     BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW, &dmc->dmc_maps[i]);
808 		if (error) {
809 			aprint_error_dev(sc->sc_dev,
810 			    "failed to creat dma map cache "
811 			    "entry %u of %zu: %d\n",
812 			    i, maxmaps, error);
813 			while (i-- > 0) {
814 				bus_dmamap_destroy(sc->sc_dmat,
815 				    dmc->dmc_maps[i]);
816 			}
817 			kmem_intr_free(dmc, dmc_size);
818 			return error;
819 		}
820 		KASSERT(dmc->dmc_maps[i] != NULL);
821 	}
822 
823 	*dmc_p = dmc;
824 
825 	return 0;
826 }
827 
828 #if 0
829 static void
830 bcmeth_dmamem_free(
831 	bus_dma_tag_t dmat,
832 	size_t map_size,
833 	bus_dma_segment_t *seg,
834 	bus_dmamap_t map,
835 	void *kvap)
836 {
837 	bus_dmamap_destroy(dmat, map);
838 	bus_dmamem_unmap(dmat, kvap, map_size);
839 	bus_dmamem_free(dmat, seg, 1);
840 }
841 #endif
842 
843 static int
844 bcmeth_dmamem_alloc(
845 	bus_dma_tag_t dmat,
846 	size_t map_size,
847 	bus_dma_segment_t *seg,
848 	bus_dmamap_t *map,
849 	void **kvap)
850 {
851 	int error;
852 	int nseg;
853 
854 	*kvap = NULL;
855 	*map = NULL;
856 
857 	error = bus_dmamem_alloc(dmat, map_size, 2*PAGE_SIZE, 0,
858 	   seg, 1, &nseg, 0);
859 	if (error)
860 		return error;
861 
862 	KASSERT(nseg == 1);
863 
864 	error = bus_dmamem_map(dmat, seg, nseg, map_size, (void **)kvap, 0);
865 	if (error == 0) {
866 		error = bus_dmamap_create(dmat, map_size, 1, map_size, 0, 0,
867 		    map);
868 		if (error == 0) {
869 			error = bus_dmamap_load(dmat, *map, *kvap, map_size,
870 			    NULL, 0);
871 			if (error == 0)
872 				return 0;
873 			bus_dmamap_destroy(dmat, *map);
874 			*map = NULL;
875 		}
876 		bus_dmamem_unmap(dmat, *kvap, map_size);
877 		*kvap = NULL;
878 	}
879 	bus_dmamem_free(dmat, seg, nseg);
880 	return 0;
881 }
882 
883 static struct mbuf *
884 bcmeth_rx_buf_alloc(
885 	struct bcmeth_softc *sc)
886 {
887 	struct mbuf *m = m_gethdr(M_DONTWAIT, MT_DATA);
888 	if (m == NULL) {
889 		printf("%s:%d: %s\n", __func__, __LINE__, "m_gethdr");
890 		return NULL;
891 	}
892 	MCLGET(m, M_DONTWAIT);
893 	if ((m->m_flags & M_EXT) == 0) {
894 		printf("%s:%d: %s\n", __func__, __LINE__, "MCLGET");
895 		m_freem(m);
896 		return NULL;
897 	}
898 	m->m_len = m->m_pkthdr.len = m->m_ext.ext_size;
899 
900 	bus_dmamap_t map = bcmeth_mapcache_get(sc, sc->sc_rx_mapcache);
901 	if (map == NULL) {
902 		printf("%s:%d: %s\n", __func__, __LINE__, "map get");
903 		m_freem(m);
904 		return NULL;
905 	}
906 	M_SETCTX(m, map);
907 	m->m_len = m->m_pkthdr.len = MCLBYTES;
908 	int error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
909 	    BUS_DMA_READ | BUS_DMA_NOWAIT);
910 	if (error) {
911 		aprint_error_dev(sc->sc_dev, "fail to load rx dmamap: %d\n",
912 		    error);
913 		M_SETCTX(m, NULL);
914 		m_freem(m);
915 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
916 		return NULL;
917 	}
918 	KASSERT(map->dm_mapsize == MCLBYTES);
919 #ifdef BCMETH_RCVMAGIC
920 	*mtod(m, uint32_t *) = htole32(BCMETH_RCVMAGIC);
921 	bus_dmamap_sync(sc->sc_dmat, map, 0, sizeof(uint32_t),
922 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
923 	bus_dmamap_sync(sc->sc_dmat, map, sizeof(uint32_t),
924 	    map->dm_mapsize - sizeof(uint32_t), BUS_DMASYNC_PREREAD);
925 #else
926 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
927 	    BUS_DMASYNC_PREREAD);
928 #endif
929 
930 	return m;
931 }
932 
933 static void
934 bcmeth_rx_map_unload(
935 	struct bcmeth_softc *sc,
936 	struct mbuf *m)
937 {
938 	KASSERT(m);
939 	for (; m != NULL; m = m->m_next) {
940 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
941 		KASSERT(map);
942 		KASSERT(map->dm_mapsize == MCLBYTES);
943 		bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_len,
944 		    BUS_DMASYNC_POSTREAD);
945 		bus_dmamap_unload(sc->sc_dmat, map);
946 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
947 		M_SETCTX(m, NULL);
948 	}
949 }
950 
951 static bool
952 bcmeth_rxq_produce(
953 	struct bcmeth_softc *sc,
954 	struct bcmeth_rxqueue *rxq)
955 {
956 	struct gmac_rxdb *producer = rxq->rxq_producer;
957 	bool produced = false;
958 
959 	while (rxq->rxq_inuse < rxq->rxq_threshold) {
960 		struct mbuf *m;
961 		IF_DEQUEUE(&sc->sc_rx_bufcache, m);
962 		if (m == NULL) {
963 			m = bcmeth_rx_buf_alloc(sc);
964 			if (m == NULL) {
965 				printf("%s: bcmeth_rx_buf_alloc failed\n",
966 				    __func__);
967 				break;
968 			}
969 		}
970 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
971 		KASSERT(map);
972 
973 		producer->rxdb_buflen = htole32(MCLBYTES);
974 		producer->rxdb_addrlo = htole32(map->dm_segs[0].ds_addr);
975 		producer->rxdb_flags &= htole32(RXDB_FLAG_ET);
976 		*rxq->rxq_mtail = m;
977 		rxq->rxq_mtail = &m->m_next;
978 		m->m_len = MCLBYTES;
979 		m->m_next = NULL;
980 		rxq->rxq_inuse++;
981 		if (++producer == rxq->rxq_last) {
982 			membar_producer();
983 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
984 			    rxq->rxq_last - rxq->rxq_producer);
985 			producer = rxq->rxq_producer = rxq->rxq_first;
986 		}
987 		produced = true;
988 	}
989 	if (produced) {
990 		membar_producer();
991 		if (producer != rxq->rxq_producer) {
992 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
993 			    producer - rxq->rxq_producer);
994 			rxq->rxq_producer = producer;
995 		}
996 		bcmeth_write_4(sc, rxq->rxq_reg_rcvptr,
997 		    rxq->rxq_descmap->dm_segs[0].ds_addr
998 		    + ((uintptr_t)producer & RCVPTR));
999 	}
1000 	return true;
1001 }
1002 
1003 static void
1004 bcmeth_rx_input(
1005 	struct bcmeth_softc *sc,
1006 	struct mbuf *m,
1007 	uint32_t rxdb_flags)
1008 {
1009 	struct ifnet * const ifp = &sc->sc_if;
1010 
1011 	bcmeth_rx_map_unload(sc, m);
1012 
1013 	m_adj(m, sc->sc_rcvoffset);
1014 
1015 	/*
1016 	 * If we are in promiscuous mode and this isn't a multicast, check the
1017 	 * destination address to make sure it matches our own.  If it doesn't,
1018 	 * mark the packet as being received promiscuously.
1019 	 */
1020 	if ((sc->sc_cmdcfg & PROMISC_EN)
1021 	    && (m->m_data[0] & 1) == 0
1022 	    && (*(uint16_t *)&m->m_data[0] != sc->sc_macaddr[0]
1023 		|| *(uint32_t *)&m->m_data[2] != sc->sc_macaddr[1])) {
1024 		m->m_flags |= M_PROMISC;
1025 	}
1026 	m_set_rcvif(m, ifp);
1027 
1028 	/*
1029 	 * Let's give it to the network subsystm to deal with.
1030 	 */
1031 #ifdef BCMETH_MPSAFE
1032 	mutex_exit(sc->sc_lock);
1033 	if_input(ifp, m);
1034 	mutex_enter(sc->sc_lock);
1035 #else
1036 	int s = splnet();
1037 	if_input(ifp, m);
1038 	splx(s);
1039 #endif
1040 }
1041 
1042 static bool
1043 bcmeth_rxq_consume(
1044 	struct bcmeth_softc *sc,
1045 	struct bcmeth_rxqueue *rxq,
1046 	size_t atmost)
1047 {
1048 	struct ifnet * const ifp = &sc->sc_if;
1049 	struct gmac_rxdb *consumer = rxq->rxq_consumer;
1050 	size_t rxconsumed = 0;
1051 	bool didconsume = false;
1052 
1053 	while (atmost-- > 0) {
1054 		if (consumer == rxq->rxq_producer) {
1055 			KASSERT(rxq->rxq_inuse == 0);
1056 			break;
1057 		}
1058 
1059 		uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1060 		uint32_t currdscr = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1061 		if (consumer == rxq->rxq_first + currdscr) {
1062 			break;
1063 		}
1064 		bcmeth_rxq_desc_postsync(sc, rxq, consumer, 1);
1065 
1066 		/*
1067 		 * We own this packet again.  Copy the rxsts word from it.
1068 		 */
1069 		rxconsumed++;
1070 		didconsume = true;
1071 		uint32_t rxsts;
1072 		KASSERT(rxq->rxq_mhead != NULL);
1073 		bus_dmamap_t map = M_GETCTX(rxq->rxq_mhead, bus_dmamap_t);
1074 		bus_dmamap_sync(sc->sc_dmat, map, 0, arm_dcache_align,
1075 		    BUS_DMASYNC_POSTREAD);
1076 		memcpy(&rxsts, rxq->rxq_mhead->m_data, 4);
1077 		rxsts = le32toh(rxsts);
1078 #if 0
1079 		KASSERTMSG(rxsts != BCMETH_RCVMAGIC, "currdscr=%u consumer=%zd",
1080 		    currdscr, consumer - rxq->rxq_first);
1081 #endif
1082 
1083 		/*
1084 		 * Get the count of descriptors.  Fetch the correct number
1085 		 * of mbufs.
1086 		 */
1087 #ifdef BCMETH_RCVMAGIC
1088 		size_t desc_count = rxsts != BCMETH_RCVMAGIC
1089 		    ? __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1 : 1;
1090 #else
1091 		size_t desc_count = __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1;
1092 #endif
1093 		struct mbuf *m = rxq->rxq_mhead;
1094 		struct mbuf *m_last = m;
1095 		for (size_t i = 1; i < desc_count; i++) {
1096 			if (++consumer == rxq->rxq_last) {
1097 				consumer = rxq->rxq_first;
1098 			}
1099 			KASSERTMSG(consumer != rxq->rxq_first + currdscr,
1100 			    "i=%zu rxsts=%#x desc_count=%zu currdscr=%u "
1101 			    "consumer=%zd", i, rxsts, desc_count, currdscr,
1102 			    consumer - rxq->rxq_first);
1103 			m_last = m_last->m_next;
1104 		}
1105 
1106 		/*
1107 		 * Now remove it/them from the list of enqueued mbufs.
1108 		 */
1109 		if ((rxq->rxq_mhead = m_last->m_next) == NULL)
1110 			rxq->rxq_mtail = &rxq->rxq_mhead;
1111 		m_last->m_next = NULL;
1112 
1113 #ifdef BCMETH_RCVMAGIC
1114 		if (rxsts == BCMETH_RCVMAGIC) {
1115 			if_statinc(ifp, if_ierrors);
1116 			if ((m->m_ext.ext_paddr >> 28) == 8) {
1117 				BCMETH_EVCNT_INCR(sc->sc_ev_rx_badmagic_lo);
1118 			} else {
1119 				BCMETH_EVCNT_INCR( sc->sc_ev_rx_badmagic_hi);
1120 			}
1121 			IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1122 		} else
1123 #endif /* BCMETH_RCVMAGIC */
1124 		if (rxsts
1125 		    & (RXSTS_CRC_ERROR |RXSTS_OVERSIZED |RXSTS_PKT_OVERFLOW)) {
1126 			aprint_error_dev(sc->sc_dev,
1127 			    "[%zu]: count=%zu rxsts=%#x\n",
1128 			    consumer - rxq->rxq_first, desc_count, rxsts);
1129 			/*
1130 			 * We encountered an error, take the mbufs and add them
1131 			 * to the rx bufcache so we can quickly reuse them.
1132 			 */
1133 			if_statinc(ifp, if_ierrors);
1134 			do {
1135 				struct mbuf *m0 = m->m_next;
1136 				m->m_next = NULL;
1137 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1138 				m = m0;
1139 			} while (m);
1140 		} else {
1141 			uint32_t framelen = __SHIFTOUT(rxsts, RXSTS_FRAMELEN);
1142 			framelen += sc->sc_rcvoffset;
1143 			m->m_pkthdr.len = framelen;
1144 			if (desc_count == 1) {
1145 				KASSERT(framelen <= MCLBYTES);
1146 				m->m_len = framelen;
1147 			} else {
1148 				m_last->m_len = framelen & (MCLBYTES - 1);
1149 			}
1150 
1151 #ifdef BCMETH_MPSAFE
1152 			/*
1153 			 * Wrap at the last entry!
1154 			 */
1155 			if (++consumer == rxq->rxq_last) {
1156 				KASSERT(consumer[-1].rxdb_flags
1157 				    & htole32(RXDB_FLAG_ET));
1158 				rxq->rxq_consumer = rxq->rxq_first;
1159 			} else {
1160 				rxq->rxq_consumer = consumer;
1161 			}
1162 			rxq->rxq_inuse -= rxconsumed;
1163 #endif /* BCMETH_MPSAFE */
1164 
1165 			/*
1166 			 * Receive the packet (which releases our lock)
1167 			 */
1168 			bcmeth_rx_input(sc, m, rxsts);
1169 
1170 #ifdef BCMETH_MPSAFE
1171 			/*
1172 			 * Since we had to give up our lock, we need to
1173 			 * refresh these.
1174 			 */
1175 			consumer = rxq->rxq_consumer;
1176 			rxconsumed = 0;
1177 			continue;
1178 #endif /* BCMETH_MPSAFE */
1179 		}
1180 
1181 		/*
1182 		 * Wrap at the last entry!
1183 		 */
1184 		if (++consumer == rxq->rxq_last) {
1185 			KASSERT(consumer[-1].rxdb_flags & htole32(RXDB_FLAG_ET));
1186 			consumer = rxq->rxq_first;
1187 		}
1188 	}
1189 
1190 	/*
1191 	 * Update queue info.
1192 	 */
1193 	rxq->rxq_consumer = consumer;
1194 	rxq->rxq_inuse -= rxconsumed;
1195 
1196 	/*
1197 	 * Did we consume anything?
1198 	 */
1199 	return didconsume;
1200 }
1201 
1202 static void
1203 bcmeth_rxq_purge(
1204 	struct bcmeth_softc *sc,
1205 	struct bcmeth_rxqueue *rxq,
1206 	bool discard)
1207 {
1208 	struct mbuf *m;
1209 
1210 	if ((m = rxq->rxq_mhead) != NULL) {
1211 		if (discard) {
1212 			bcmeth_rx_map_unload(sc, m);
1213 			m_freem(m);
1214 		} else {
1215 			while (m != NULL) {
1216 				struct mbuf *m0 = m->m_next;
1217 				m->m_next = NULL;
1218 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1219 				m = m0;
1220 			}
1221 		}
1222 	}
1223 
1224 	rxq->rxq_mhead = NULL;
1225 	rxq->rxq_mtail = &rxq->rxq_mhead;
1226 	rxq->rxq_inuse = 0;
1227 }
1228 
1229 static void
1230 bcmeth_rxq_reset(
1231 	struct bcmeth_softc *sc,
1232 	struct bcmeth_rxqueue *rxq)
1233 {
1234 	/*
1235 	 * sync all the descriptors
1236 	 */
1237 	bcmeth_rxq_desc_postsync(sc, rxq, rxq->rxq_first,
1238 	    rxq->rxq_last - rxq->rxq_first);
1239 
1240 	/*
1241 	 * Make sure we own all descriptors in the ring.
1242 	 */
1243 	struct gmac_rxdb *rxdb;
1244 	for (rxdb = rxq->rxq_first; rxdb < rxq->rxq_last - 1; rxdb++) {
1245 		rxdb->rxdb_flags = htole32(RXDB_FLAG_IC);
1246 	}
1247 
1248 	/*
1249 	 * Last descriptor has the wrap flag.
1250 	 */
1251 	rxdb->rxdb_flags = htole32(RXDB_FLAG_ET | RXDB_FLAG_IC);
1252 
1253 	/*
1254 	 * Reset the producer consumer indexes.
1255 	 */
1256 	rxq->rxq_consumer = rxq->rxq_first;
1257 	rxq->rxq_producer = rxq->rxq_first;
1258 	rxq->rxq_inuse = 0;
1259 	if (rxq->rxq_threshold < BCMETH_MINRXMBUFS)
1260 		rxq->rxq_threshold = BCMETH_MINRXMBUFS;
1261 
1262 	sc->sc_intmask |= RCVINT | RCVFIFOOF | RCVDESCUF;
1263 
1264 	/*
1265 	 * Restart the receiver at the first descriptor
1266 	 */
1267 	bcmeth_write_4(sc, rxq->rxq_reg_rcvaddrlo,
1268 	    rxq->rxq_descmap->dm_segs[0].ds_addr);
1269 }
1270 
1271 static int
1272 bcmeth_rxq_attach(
1273 	struct bcmeth_softc *sc,
1274 	struct bcmeth_rxqueue *rxq,
1275 	u_int qno)
1276 {
1277 	size_t desc_count = BCMETH_RINGSIZE / sizeof(rxq->rxq_first[0]);
1278 	int error;
1279 	void *descs;
1280 
1281 	KASSERT(desc_count == 256 || desc_count == 512);
1282 
1283 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1284 	   &rxq->rxq_descmap_seg, &rxq->rxq_descmap, &descs);
1285 	if (error)
1286 		return error;
1287 
1288 	memset(descs, 0, BCMETH_RINGSIZE);
1289 	rxq->rxq_first = descs;
1290 	rxq->rxq_last = rxq->rxq_first + desc_count;
1291 	rxq->rxq_consumer = descs;
1292 	rxq->rxq_producer = descs;
1293 
1294 	bcmeth_rxq_purge(sc, rxq, true);
1295 	bcmeth_rxq_reset(sc, rxq);
1296 
1297 	rxq->rxq_reg_rcvaddrlo = GMAC_RCVADDR_LOW;
1298 	rxq->rxq_reg_rcvctl = GMAC_RCVCONTROL;
1299 	rxq->rxq_reg_rcvptr = GMAC_RCVPTR;
1300 	rxq->rxq_reg_rcvsts0 = GMAC_RCVSTATUS0;
1301 	rxq->rxq_reg_rcvsts1 = GMAC_RCVSTATUS1;
1302 
1303 	return 0;
1304 }
1305 
1306 static bool
1307 bcmeth_txq_active_p(
1308 	struct bcmeth_softc * const sc,
1309 	struct bcmeth_txqueue *txq)
1310 {
1311 	return !IF_IS_EMPTY(&txq->txq_mbufs);
1312 }
1313 
1314 static bool
1315 bcmeth_txq_fillable_p(
1316 	struct bcmeth_softc * const sc,
1317 	struct bcmeth_txqueue *txq)
1318 {
1319 	return txq->txq_free >= txq->txq_threshold;
1320 }
1321 
1322 static int
1323 bcmeth_txq_attach(
1324 	struct bcmeth_softc *sc,
1325 	struct bcmeth_txqueue *txq,
1326 	u_int qno)
1327 {
1328 	size_t desc_count = BCMETH_RINGSIZE / sizeof(txq->txq_first[0]);
1329 	int error;
1330 	void *descs;
1331 
1332 	KASSERT(desc_count == 256 || desc_count == 512);
1333 
1334 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1335 	   &txq->txq_descmap_seg, &txq->txq_descmap, &descs);
1336 	if (error)
1337 		return error;
1338 
1339 	memset(descs, 0, BCMETH_RINGSIZE);
1340 	txq->txq_first = descs;
1341 	txq->txq_last = txq->txq_first + desc_count;
1342 	txq->txq_consumer = descs;
1343 	txq->txq_producer = descs;
1344 
1345 	IFQ_SET_MAXLEN(&txq->txq_mbufs, BCMETH_MAXTXMBUFS);
1346 
1347 	txq->txq_reg_xmtaddrlo = GMAC_XMTADDR_LOW;
1348 	txq->txq_reg_xmtctl = GMAC_XMTCONTROL;
1349 	txq->txq_reg_xmtptr = GMAC_XMTPTR;
1350 	txq->txq_reg_xmtsts0 = GMAC_XMTSTATUS0;
1351 	txq->txq_reg_xmtsts1 = GMAC_XMTSTATUS1;
1352 
1353 	bcmeth_txq_reset(sc, txq);
1354 
1355 	return 0;
1356 }
1357 
1358 static int
1359 bcmeth_txq_map_load(
1360 	struct bcmeth_softc *sc,
1361 	struct bcmeth_txqueue *txq,
1362 	struct mbuf *m)
1363 {
1364 	bus_dmamap_t map;
1365 	int error;
1366 
1367 	map = M_GETCTX(m, bus_dmamap_t);
1368 	if (map != NULL)
1369 		return 0;
1370 
1371 	map = bcmeth_mapcache_get(sc, sc->sc_tx_mapcache);
1372 	if (map == NULL)
1373 		return ENOMEM;
1374 
1375 	error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
1376 	    BUS_DMA_WRITE | BUS_DMA_NOWAIT);
1377 	if (error)
1378 		return error;
1379 
1380 	bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_pkthdr.len,
1381 	    BUS_DMASYNC_PREWRITE);
1382 	M_SETCTX(m, map);
1383 	return 0;
1384 }
1385 
1386 static void
1387 bcmeth_txq_map_unload(
1388 	struct bcmeth_softc *sc,
1389 	struct bcmeth_txqueue *txq,
1390 	struct mbuf *m)
1391 {
1392 	KASSERT(m);
1393 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1394 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1395 	    BUS_DMASYNC_POSTWRITE);
1396 	bus_dmamap_unload(sc->sc_dmat, map);
1397 	bcmeth_mapcache_put(sc, sc->sc_tx_mapcache, map);
1398 }
1399 
1400 static bool
1401 bcmeth_txq_produce(
1402 	struct bcmeth_softc *sc,
1403 	struct bcmeth_txqueue *txq,
1404 	struct mbuf *m)
1405 {
1406 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1407 
1408 	if (map->dm_nsegs > txq->txq_free)
1409 		return false;
1410 
1411 	/*
1412 	 * TCP Offload flag must be set in the first descriptor.
1413 	 */
1414 	struct gmac_txdb *producer = txq->txq_producer;
1415 	uint32_t first_flags = TXDB_FLAG_SF;
1416 	uint32_t last_flags = TXDB_FLAG_EF;
1417 
1418 	/*
1419 	 * If we've produced enough descriptors without consuming any
1420 	 * we need to ask for an interrupt to reclaim some.
1421 	 */
1422 	txq->txq_lastintr += map->dm_nsegs;
1423 	if (txq->txq_lastintr >= txq->txq_threshold
1424 	    || txq->txq_mbufs.ifq_len + 1 == txq->txq_mbufs.ifq_maxlen) {
1425 		txq->txq_lastintr = 0;
1426 		last_flags |= TXDB_FLAG_IC;
1427 	}
1428 
1429 	KASSERT(producer != txq->txq_last);
1430 
1431 	struct gmac_txdb *start = producer;
1432 	size_t count = map->dm_nsegs;
1433 	producer->txdb_flags |= htole32(first_flags);
1434 	producer->txdb_addrlo = htole32(map->dm_segs[0].ds_addr);
1435 	producer->txdb_buflen = htole32(map->dm_segs[0].ds_len);
1436 	for (u_int i = 1; i < map->dm_nsegs; i++) {
1437 #if 0
1438 		printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1439 		    le32toh(producer->txdb_flags),
1440 		    le32toh(producer->txdb_buflen),
1441 		    le32toh(producer->txdb_addrlo),
1442 		    le32toh(producer->txdb_addrhi));
1443 #endif
1444 		if (__predict_false(++producer == txq->txq_last)) {
1445 			bcmeth_txq_desc_presync(sc, txq, start,
1446 			    txq->txq_last - start);
1447 			count -= txq->txq_last - start;
1448 			producer = txq->txq_first;
1449 			start = txq->txq_first;
1450 		}
1451 		producer->txdb_addrlo = htole32(map->dm_segs[i].ds_addr);
1452 		producer->txdb_buflen = htole32(map->dm_segs[i].ds_len);
1453 	}
1454 	producer->txdb_flags |= htole32(last_flags);
1455 #if 0
1456 	printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1457 	    le32toh(producer->txdb_flags), le32toh(producer->txdb_buflen),
1458 	    le32toh(producer->txdb_addrlo), le32toh(producer->txdb_addrhi));
1459 #endif
1460 	if (count)
1461 		bcmeth_txq_desc_presync(sc, txq, start, count);
1462 
1463 	/*
1464 	 * Reduce free count by the number of segments we consumed.
1465 	 */
1466 	txq->txq_free -= map->dm_nsegs;
1467 	KASSERT(map->dm_nsegs == 1 || txq->txq_producer != producer);
1468 	KASSERT(map->dm_nsegs == 1
1469 	    || (txq->txq_producer->txdb_flags & htole32(TXDB_FLAG_EF)) == 0);
1470 	KASSERT(producer->txdb_flags & htole32(TXDB_FLAG_EF));
1471 
1472 #if 0
1473 	printf("%s: mbuf %p: produced a %u byte packet in %u segments "
1474 	    "(%zd..%zd)\n", __func__, m, m->m_pkthdr.len, map->dm_nsegs,
1475 	    txq->txq_producer - txq->txq_first, producer - txq->txq_first);
1476 #endif
1477 
1478 	if (producer + 1 == txq->txq_last)
1479 		txq->txq_producer = txq->txq_first;
1480 	else
1481 		txq->txq_producer = producer + 1;
1482 	IF_ENQUEUE(&txq->txq_mbufs, m);
1483 
1484 	/*
1485 	 * Let the transmitter know there's more to do
1486 	 */
1487 	bcmeth_write_4(sc, txq->txq_reg_xmtptr,
1488 	    txq->txq_descmap->dm_segs[0].ds_addr
1489 	    + ((uintptr_t)txq->txq_producer & XMT_LASTDSCR));
1490 
1491 	return true;
1492 }
1493 
1494 static struct mbuf *
1495 bcmeth_copy_packet(struct mbuf *m)
1496 {
1497 	struct mbuf *mext = NULL;
1498 	size_t misalignment = 0;
1499 	size_t hlen = 0;
1500 
1501 	for (mext = m; mext != NULL; mext = mext->m_next) {
1502 		if (mext->m_flags & M_EXT) {
1503 			misalignment = mtod(mext, vaddr_t) & arm_dcache_align;
1504 			break;
1505 		}
1506 		hlen += m->m_len;
1507 	}
1508 
1509 	struct mbuf *n = m->m_next;
1510 	if (m != mext && hlen + misalignment <= MHLEN && false) {
1511 		KASSERT(m->m_pktdat <= m->m_data
1512 		    && m->m_data <= &m->m_pktdat[MHLEN - m->m_len]);
1513 		size_t oldoff = m->m_data - m->m_pktdat;
1514 		size_t off;
1515 		if (mext == NULL) {
1516 			off = (oldoff + hlen > MHLEN) ? 0 : oldoff;
1517 		} else {
1518 			off = MHLEN - (hlen + misalignment);
1519 		}
1520 		KASSERT(off + hlen + misalignment <= MHLEN);
1521 		if (((oldoff ^ off) & arm_dcache_align) != 0 || off < oldoff) {
1522 			memmove(&m->m_pktdat[off], m->m_data, m->m_len);
1523 			m->m_data = &m->m_pktdat[off];
1524 		}
1525 		m_copydata(n, 0, hlen - m->m_len, &m->m_data[m->m_len]);
1526 		m->m_len = hlen;
1527 		m->m_next = mext;
1528 		while (n != mext) {
1529 			n = m_free(n);
1530 		}
1531 		return m;
1532 	}
1533 
1534 	struct mbuf *m0 = m_gethdr(M_DONTWAIT, m->m_type);
1535 	if (m0 == NULL) {
1536 		return NULL;
1537 	}
1538 	m_copy_pkthdr(m0, m);
1539 	MCLAIM(m0, m->m_owner);
1540 	if (m0->m_pkthdr.len > MHLEN) {
1541 		MCLGET(m0, M_DONTWAIT);
1542 		if ((m0->m_flags & M_EXT) == 0) {
1543 			m_freem(m0);
1544 			return NULL;
1545 		}
1546 	}
1547 	m0->m_len = m->m_pkthdr.len;
1548 	m_copydata(m, 0, m0->m_len, mtod(m0, void *));
1549 	m_freem(m);
1550 	return m0;
1551 }
1552 
1553 static bool
1554 bcmeth_txq_enqueue(
1555 	struct bcmeth_softc *sc,
1556 	struct bcmeth_txqueue *txq)
1557 {
1558 	for (;;) {
1559 		if (IF_QFULL(&txq->txq_mbufs))
1560 			return false;
1561 		struct mbuf *m = txq->txq_next;
1562 		if (m == NULL) {
1563 			int s = splnet();
1564 			IF_DEQUEUE(&sc->sc_if.if_snd, m);
1565 			splx(s);
1566 			if (m == NULL)
1567 				return true;
1568 			M_SETCTX(m, NULL);
1569 		} else {
1570 			txq->txq_next = NULL;
1571 		}
1572 		/*
1573 		 * If LINK2 is set and this packet uses multiple mbufs,
1574 		 * consolidate it into a single mbuf.
1575 		 */
1576 		if (m->m_next != NULL && (sc->sc_if.if_flags & IFF_LINK2)) {
1577 			struct mbuf *m0 = bcmeth_copy_packet(m);
1578 			if (m0 == NULL) {
1579 				txq->txq_next = m;
1580 				return true;
1581 			}
1582 			m = m0;
1583 		}
1584 		int error = bcmeth_txq_map_load(sc, txq, m);
1585 		if (error) {
1586 			aprint_error_dev(sc->sc_dev,
1587 			    "discarded packet due to "
1588 			    "dmamap load failure: %d\n", error);
1589 			m_freem(m);
1590 			continue;
1591 		}
1592 		KASSERT(txq->txq_next == NULL);
1593 		if (!bcmeth_txq_produce(sc, txq, m)) {
1594 			txq->txq_next = m;
1595 			return false;
1596 		}
1597 		KASSERT(txq->txq_next == NULL);
1598 	}
1599 }
1600 
1601 static bool
1602 bcmeth_txq_consume(
1603 	struct bcmeth_softc *sc,
1604 	struct bcmeth_txqueue *txq)
1605 {
1606 	struct ifnet * const ifp = &sc->sc_if;
1607 	struct gmac_txdb *consumer = txq->txq_consumer;
1608 	size_t txfree = 0;
1609 
1610 #if 0
1611 	printf("%s: entry: free=%zu\n", __func__, txq->txq_free);
1612 #endif
1613 
1614 	for (;;) {
1615 		if (consumer == txq->txq_producer) {
1616 			txq->txq_consumer = consumer;
1617 			txq->txq_free += txfree;
1618 			txq->txq_lastintr -= uimin(txq->txq_lastintr, txfree);
1619 #if 0
1620 			printf("%s: empty: freed %zu descriptors going from "
1621 			    "%zu to %zu\n", __func__, txfree,
1622 			    txq->txq_free - txfree, txq->txq_free);
1623 #endif
1624 			KASSERT(txq->txq_lastintr == 0);
1625 			KASSERT(txq->txq_free
1626 			    == txq->txq_last - txq->txq_first - 1);
1627 			return true;
1628 		}
1629 		bcmeth_txq_desc_postsync(sc, txq, consumer, 1);
1630 		uint32_t s0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
1631 		if (consumer == txq->txq_first + __SHIFTOUT(s0, XMT_CURRDSCR)) {
1632 			txq->txq_consumer = consumer;
1633 			txq->txq_free += txfree;
1634 			txq->txq_lastintr -= uimin(txq->txq_lastintr, txfree);
1635 #if 0
1636 			printf("%s: freed %zu descriptors\n",
1637 			    __func__, txfree);
1638 #endif
1639 			return bcmeth_txq_fillable_p(sc, txq);
1640 		}
1641 
1642 		/*
1643 		 * If this is the last descriptor in the chain, get the
1644 		 * mbuf, free its dmamap, and free the mbuf chain itself.
1645 		 */
1646 		const uint32_t txdb_flags = le32toh(consumer->txdb_flags);
1647 		if (txdb_flags & TXDB_FLAG_EF) {
1648 			struct mbuf *m;
1649 
1650 			IF_DEQUEUE(&txq->txq_mbufs, m);
1651 			KASSERT(m);
1652 			bcmeth_txq_map_unload(sc, txq, m);
1653 #if 0
1654 			printf("%s: mbuf %p: consumed a %u byte packet\n",
1655 			    __func__, m, m->m_pkthdr.len);
1656 #endif
1657 			bpf_mtap(ifp, m, BPF_D_OUT);
1658 			if_statinc(ifp, if_opackets);
1659 			if_statadd(ifp, if_obytes,  m->m_pkthdr.len);
1660 			if (m->m_flags & M_MCAST)
1661 				if_statinc(ifp, if_omcasts);
1662 			m_freem(m);
1663 		}
1664 
1665 		/*
1666 		 * We own this packet again.  Clear all flags except wrap.
1667 		 */
1668 		txfree++;
1669 
1670 		/*
1671 		 * Wrap at the last entry!
1672 		 */
1673 		if (txdb_flags & TXDB_FLAG_ET) {
1674 			consumer->txdb_flags = htole32(TXDB_FLAG_ET);
1675 			KASSERT(consumer + 1 == txq->txq_last);
1676 			consumer = txq->txq_first;
1677 		} else {
1678 			consumer->txdb_flags = 0;
1679 			consumer++;
1680 			KASSERT(consumer < txq->txq_last);
1681 		}
1682 	}
1683 }
1684 
1685 static void
1686 bcmeth_txq_purge(
1687 	struct bcmeth_softc *sc,
1688 	struct bcmeth_txqueue *txq)
1689 {
1690 	struct mbuf *m;
1691 	KASSERT((bcmeth_read_4(sc, UNIMAC_COMMAND_CONFIG) & TX_ENA) == 0);
1692 
1693 	for (;;) {
1694 		IF_DEQUEUE(&txq->txq_mbufs, m);
1695 		if (m == NULL)
1696 			break;
1697 		bcmeth_txq_map_unload(sc, txq, m);
1698 		m_freem(m);
1699 	}
1700 	if ((m = txq->txq_next) != NULL) {
1701 		txq->txq_next = NULL;
1702 		bcmeth_txq_map_unload(sc, txq, m);
1703 		m_freem(m);
1704 	}
1705 }
1706 
1707 static void
1708 bcmeth_txq_reset(
1709 	struct bcmeth_softc *sc,
1710 	struct bcmeth_txqueue *txq)
1711 {
1712 	/*
1713 	 * sync all the descriptors
1714 	 */
1715 	bcmeth_txq_desc_postsync(sc, txq, txq->txq_first,
1716 	    txq->txq_last - txq->txq_first);
1717 
1718 	/*
1719 	 * Make sure we own all descriptors in the ring.
1720 	 */
1721 	struct gmac_txdb *txdb;
1722 	for (txdb = txq->txq_first; txdb < txq->txq_last - 1; txdb++) {
1723 		txdb->txdb_flags = 0;
1724 	}
1725 
1726 	/*
1727 	 * Last descriptor has the wrap flag.
1728 	 */
1729 	txdb->txdb_flags = htole32(TXDB_FLAG_ET);
1730 
1731 	/*
1732 	 * Reset the producer consumer indexes.
1733 	 */
1734 	txq->txq_consumer = txq->txq_first;
1735 	txq->txq_producer = txq->txq_first;
1736 	txq->txq_free = txq->txq_last - txq->txq_first - 1;
1737 	txq->txq_threshold = txq->txq_free / 2;
1738 	txq->txq_lastintr = 0;
1739 
1740 	/*
1741 	 * What do we want to get interrupted on?
1742 	 */
1743 	sc->sc_intmask |= XMTINT_0 | XMTUF;
1744 
1745 	/*
1746 	 * Restart the transmiter at the first descriptor
1747 	 */
1748 	bcmeth_write_4(sc, txq->txq_reg_xmtaddrlo,
1749 	    txq->txq_descmap->dm_segs->ds_addr);
1750 }
1751 
1752 static void
1753 bcmeth_ifstart(struct ifnet *ifp)
1754 {
1755 	struct bcmeth_softc * const sc = ifp->if_softc;
1756 
1757 	if (__predict_false((ifp->if_flags & IFF_RUNNING) == 0)) {
1758 		return;
1759 	}
1760 
1761 #ifdef BCMETH_MPSAFETX
1762 	if (cpu_intr_p()) {
1763 #endif
1764 		atomic_or_uint(&sc->sc_soft_flags, SOFT_TXINTR);
1765 		softint_schedule(sc->sc_soft_ih);
1766 #ifdef BCMETH_MPSAFETX
1767 	} else {
1768 		/*
1769 		 * Either we are in a softintr thread already or some other
1770 		 * thread so just borrow it to do the send and save ourselves
1771 		 * the overhead of a fast soft int.
1772 		 */
1773 		bcmeth_soft_txintr(sc);
1774 	}
1775 #endif
1776 }
1777 
1778 int
1779 bcmeth_intr(void *arg)
1780 {
1781 	struct bcmeth_softc * const sc = arg;
1782 	uint32_t soft_flags = 0;
1783 	uint32_t work_flags = 0;
1784 	int rv = 0;
1785 
1786 	mutex_enter(sc->sc_hwlock);
1787 
1788 	uint32_t intmask = sc->sc_intmask;
1789 	BCMETH_EVCNT_INCR(sc->sc_ev_intr);
1790 
1791 	for (;;) {
1792 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
1793 		intstatus &= intmask;
1794 		bcmeth_write_4(sc, GMAC_INTSTATUS, intstatus);	/* write 1 to clear */
1795 		if (intstatus == 0) {
1796 			break;
1797 		}
1798 #if 0
1799 		aprint_normal_dev(sc->sc_dev, "%s: intstatus=%#x intmask=%#x\n",
1800 		    __func__, intstatus, bcmeth_read_4(sc, GMAC_INTMASK));
1801 #endif
1802 		if (intstatus & RCVINT) {
1803 			struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
1804 			intmask &= ~RCVINT;
1805 
1806 			uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1807 			uint32_t descs = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1808 			if (descs < rxq->rxq_consumer - rxq->rxq_first) {
1809 				/*
1810 				 * We wrapped at the end so count how far
1811 				 * we are from the end.
1812 				 */
1813 				descs += rxq->rxq_last - rxq->rxq_consumer;
1814 			} else {
1815 				descs -= rxq->rxq_consumer - rxq->rxq_first;
1816 			}
1817 			/*
1818 			 * If we "timedout" we can't be hogging so use
1819 			 * softints.  If we exceeded then we might hogging
1820 			 * so let the workqueue deal with them.
1821 			 */
1822 			const uint32_t framecount = __SHIFTOUT(sc->sc_rcvlazy,
1823 			    INTRCVLAZY_FRAMECOUNT);
1824 			if (descs < framecount
1825 			    || (curcpu()->ci_curlwp->l_flag & LW_IDLE)) {
1826 				soft_flags |= SOFT_RXINTR;
1827 			} else {
1828 				work_flags |= WORK_RXINTR;
1829 			}
1830 		}
1831 
1832 		if (intstatus & XMTINT_0) {
1833 			intmask &= ~XMTINT_0;
1834 			soft_flags |= SOFT_TXINTR;
1835 		}
1836 
1837 		if (intstatus & RCVDESCUF) {
1838 			intmask &= ~RCVDESCUF;
1839 			work_flags |= WORK_RXUNDERFLOW;
1840 		}
1841 
1842 		intstatus &= intmask;
1843 		if (intstatus) {
1844 			aprint_error_dev(sc->sc_dev,
1845 			    "intr: intstatus=%#x\n", intstatus);
1846 			aprint_error_dev(sc->sc_dev,
1847 			    "rcvbase=%p/%#lx rcvptr=%#x rcvsts=%#x/%#x\n",
1848 			    sc->sc_rxq.rxq_first,
1849 			    sc->sc_rxq.rxq_descmap->dm_segs[0].ds_addr,
1850 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvptr),
1851 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts0),
1852 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts1));
1853 			aprint_error_dev(sc->sc_dev,
1854 			    "xmtbase=%p/%#lx xmtptr=%#x xmtsts=%#x/%#x\n",
1855 			    sc->sc_txq.txq_first,
1856 			    sc->sc_txq.txq_descmap->dm_segs[0].ds_addr,
1857 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtptr),
1858 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts0),
1859 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts1));
1860 			intmask &= ~intstatus;
1861 			work_flags |= WORK_REINIT;
1862 			break;
1863 		}
1864 	}
1865 
1866 	if (intmask != sc->sc_intmask) {
1867 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1868 	}
1869 
1870 	if (work_flags) {
1871 		if (sc->sc_work_flags == 0) {
1872 			workqueue_enqueue(sc->sc_workq, &sc->sc_work, NULL);
1873 		}
1874 		atomic_or_32(&sc->sc_work_flags, work_flags);
1875 		rv = 1;
1876 	}
1877 
1878 	if (soft_flags) {
1879 		if (sc->sc_soft_flags == 0) {
1880 			softint_schedule(sc->sc_soft_ih);
1881 		}
1882 		atomic_or_32(&sc->sc_soft_flags, soft_flags);
1883 		rv = 1;
1884 	}
1885 
1886 	mutex_exit(sc->sc_hwlock);
1887 
1888 	return rv;
1889 }
1890 
1891 #ifdef BCMETH_MPSAFETX
1892 void
1893 bcmeth_soft_txintr(struct bcmeth_softc *sc)
1894 {
1895 	mutex_enter(sc->sc_lock);
1896 	/*
1897 	 * Let's do what we came here for.  Consume transmitted
1898 	 * packets off the transmit ring.
1899 	 */
1900 	if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1901 	    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1902 		BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1903 	}
1904 	if (sc->sc_if.if_flags & IFF_RUNNING) {
1905 		mutex_spin_enter(sc->sc_hwlock);
1906 		sc->sc_intmask |= XMTINT_0;
1907 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1908 		mutex_spin_exit(sc->sc_hwlock);
1909 	}
1910 	mutex_exit(sc->sc_lock);
1911 }
1912 #endif /* BCMETH_MPSAFETX */
1913 
1914 void
1915 bcmeth_soft_intr(void *arg)
1916 {
1917 	struct bcmeth_softc * const sc = arg;
1918 	struct ifnet * const ifp = &sc->sc_if;
1919 	uint32_t intmask = 0;
1920 
1921 	mutex_enter(sc->sc_lock);
1922 
1923 	u_int soft_flags = atomic_swap_uint(&sc->sc_soft_flags, 0);
1924 
1925 	BCMETH_EVCNT_INCR(sc->sc_ev_soft_intr);
1926 
1927 	if ((soft_flags & SOFT_TXINTR)
1928 	    || bcmeth_txq_active_p(sc, &sc->sc_txq)) {
1929 		/*
1930 		 * Let's do what we came here for.  Consume transmitted
1931 		 * packets off the transmit ring.
1932 		 */
1933 		if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1934 		    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1935 			BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1936 		}
1937 		intmask |= XMTINT_0;
1938 	}
1939 
1940 	if (soft_flags & SOFT_RXINTR) {
1941 		/*
1942 		 * Let's consume
1943 		 */
1944 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
1945 		    sc->sc_rxq.rxq_threshold / 4)) {
1946 			/*
1947 			 * We've consumed a quarter of the ring and still have
1948 			 * more to do.  Refill the ring.
1949 			 */
1950 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
1951 		}
1952 		intmask |= RCVINT;
1953 	}
1954 
1955 	if (ifp->if_flags & IFF_RUNNING) {
1956 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
1957 		mutex_spin_enter(sc->sc_hwlock);
1958 		sc->sc_intmask |= intmask;
1959 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1960 		mutex_spin_exit(sc->sc_hwlock);
1961 	}
1962 
1963 	mutex_exit(sc->sc_lock);
1964 }
1965 
1966 void
1967 bcmeth_worker(struct work *wk, void *arg)
1968 {
1969 	struct bcmeth_softc * const sc = arg;
1970 	struct ifnet * const ifp = &sc->sc_if;
1971 	uint32_t intmask = 0;
1972 
1973 	mutex_enter(sc->sc_lock);
1974 
1975 	BCMETH_EVCNT_INCR(sc->sc_ev_work);
1976 
1977 	uint32_t work_flags = atomic_swap_32(&sc->sc_work_flags, 0);
1978 	if (work_flags & WORK_REINIT) {
1979 		int s = splnet();
1980 		sc->sc_soft_flags = 0;
1981 		bcmeth_ifinit(ifp);
1982 		splx(s);
1983 		work_flags &= ~WORK_RXUNDERFLOW;
1984 	}
1985 
1986 	if (work_flags & WORK_RXUNDERFLOW) {
1987 		struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
1988 		size_t threshold = 5 * rxq->rxq_threshold / 4;
1989 		if (threshold >= rxq->rxq_last - rxq->rxq_first) {
1990 			threshold = rxq->rxq_last - rxq->rxq_first - 1;
1991 		} else {
1992 			intmask |= RCVDESCUF;
1993 		}
1994 		aprint_normal_dev(sc->sc_dev,
1995 		    "increasing receive buffers from %zu to %zu\n",
1996 		    rxq->rxq_threshold, threshold);
1997 		rxq->rxq_threshold = threshold;
1998 	}
1999 
2000 	if (work_flags & WORK_RXINTR) {
2001 		/*
2002 		 * Let's consume
2003 		 */
2004 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
2005 		    sc->sc_rxq.rxq_threshold / 4)) {
2006 			/*
2007 			 * We've consumed a quarter of the ring and still have
2008 			 * more to do.  Refill the ring.
2009 			 */
2010 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
2011 		}
2012 		intmask |= RCVINT;
2013 	}
2014 
2015 	if (ifp->if_flags & IFF_RUNNING) {
2016 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
2017 #if 0
2018 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
2019 		if (intstatus & RCVINT) {
2020 			bcmeth_write_4(sc, GMAC_INTSTATUS, RCVINT);
2021 			work_flags |= WORK_RXINTR;
2022 			continue;
2023 		}
2024 #endif
2025 		mutex_spin_enter(sc->sc_hwlock);
2026 		sc->sc_intmask |= intmask;
2027 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
2028 		mutex_spin_exit(sc->sc_hwlock);
2029 	}
2030 
2031 	mutex_exit(sc->sc_lock);
2032 }
2033