xref: /netbsd-src/sys/arch/arm/broadcom/bcm53xx_eth.c (revision c38e7cc395b1472a774ff828e46123de44c628e9)
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 1.30 2017/10/23 09:23:25 msaitoh 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 
56 #include <net/if_dl.h>
57 
58 #include <net/bpf.h>
59 
60 #include <dev/mii/miivar.h>
61 
62 #include <arm/locore.h>
63 
64 #include <arm/broadcom/bcm53xx_reg.h>
65 #include <arm/broadcom/bcm53xx_var.h>
66 
67 //#define BCMETH_MPSAFE
68 
69 #ifdef BCMETH_COUNTERS
70 #define	BCMETH_EVCNT_ADD(a,b)	((void)((a).ev_count += (b)))
71 #else
72 #define	BCMETH_EVCNT_ADD(a,b)	do { } while (/*CONSTCOND*/0)
73 #endif
74 #define	BCMETH_EVCNT_INCR(a)	BCMETH_EVCNT_ADD((a), 1)
75 
76 #define	BCMETH_MAXTXMBUFS	128
77 #define	BCMETH_NTXSEGS		30
78 #define	BCMETH_MAXRXMBUFS	255
79 #define	BCMETH_MINRXMBUFS	64
80 #define	BCMETH_NRXSEGS		1
81 #define	BCMETH_RINGSIZE		PAGE_SIZE
82 
83 #if 1
84 #define	BCMETH_RCVMAGIC		0xfeedface
85 #endif
86 
87 static int bcmeth_ccb_match(device_t, cfdata_t, void *);
88 static void bcmeth_ccb_attach(device_t, device_t, void *);
89 
90 struct bcmeth_txqueue {
91 	bus_dmamap_t txq_descmap;
92 	struct gmac_txdb *txq_consumer;
93 	struct gmac_txdb *txq_producer;
94 	struct gmac_txdb *txq_first;
95 	struct gmac_txdb *txq_last;
96 	struct ifqueue txq_mbufs;
97 	struct mbuf *txq_next;
98 	size_t txq_free;
99 	size_t txq_threshold;
100 	size_t txq_lastintr;
101 	bus_size_t txq_reg_xmtaddrlo;
102 	bus_size_t txq_reg_xmtptr;
103 	bus_size_t txq_reg_xmtctl;
104 	bus_size_t txq_reg_xmtsts0;
105 	bus_size_t txq_reg_xmtsts1;
106 	bus_dma_segment_t txq_descmap_seg;
107 };
108 
109 struct bcmeth_rxqueue {
110 	bus_dmamap_t rxq_descmap;
111 	struct gmac_rxdb *rxq_consumer;
112 	struct gmac_rxdb *rxq_producer;
113 	struct gmac_rxdb *rxq_first;
114 	struct gmac_rxdb *rxq_last;
115 	struct mbuf *rxq_mhead;
116 	struct mbuf **rxq_mtail;
117 	struct mbuf *rxq_mconsumer;
118 	size_t rxq_inuse;
119 	size_t rxq_threshold;
120 	bus_size_t rxq_reg_rcvaddrlo;
121 	bus_size_t rxq_reg_rcvptr;
122 	bus_size_t rxq_reg_rcvctl;
123 	bus_size_t rxq_reg_rcvsts0;
124 	bus_size_t rxq_reg_rcvsts1;
125 	bus_dma_segment_t rxq_descmap_seg;
126 };
127 
128 struct bcmeth_mapcache {
129 	u_int dmc_nmaps;
130 	u_int dmc_maxseg;
131 	u_int dmc_maxmaps;
132 	u_int dmc_maxmapsize;
133 	bus_dmamap_t dmc_maps[0];
134 };
135 
136 struct bcmeth_softc {
137 	device_t sc_dev;
138 	bus_space_tag_t sc_bst;
139 	bus_space_handle_t sc_bsh;
140 	bus_dma_tag_t sc_dmat;
141 	kmutex_t *sc_lock;
142 	kmutex_t *sc_hwlock;
143 	struct ethercom sc_ec;
144 #define	sc_if		sc_ec.ec_if
145 	struct ifmedia sc_media;
146 	void *sc_soft_ih;
147 	void *sc_ih;
148 
149 	struct bcmeth_rxqueue sc_rxq;
150 	struct bcmeth_txqueue sc_txq;
151 
152 	size_t sc_rcvoffset;
153 	uint32_t sc_macaddr[2];
154 	uint32_t sc_maxfrm;
155 	uint32_t sc_cmdcfg;
156 	uint32_t sc_intmask;
157 	uint32_t sc_rcvlazy;
158 	volatile uint32_t sc_soft_flags;
159 #define	SOFT_RXINTR		0x01
160 #define	SOFT_TXINTR		0x02
161 
162 #ifdef BCMETH_COUNTERS
163 	struct evcnt sc_ev_intr;
164 	struct evcnt sc_ev_soft_intr;
165 	struct evcnt sc_ev_work;
166 	struct evcnt sc_ev_tx_stall;
167 	struct evcnt sc_ev_rx_badmagic_lo;
168 	struct evcnt sc_ev_rx_badmagic_hi;
169 #endif
170 
171 	struct ifqueue sc_rx_bufcache;
172 	struct bcmeth_mapcache *sc_rx_mapcache;
173 	struct bcmeth_mapcache *sc_tx_mapcache;
174 
175 	struct workqueue *sc_workq;
176 	struct work sc_work;
177 
178 	volatile uint32_t sc_work_flags;
179 #define	WORK_RXINTR		0x01
180 #define	WORK_RXUNDERFLOW	0x02
181 #define	WORK_REINIT		0x04
182 
183 	uint8_t sc_enaddr[ETHER_ADDR_LEN];
184 };
185 
186 static void bcmeth_ifstart(struct ifnet *);
187 static void bcmeth_ifwatchdog(struct ifnet *);
188 static int bcmeth_ifinit(struct ifnet *);
189 static void bcmeth_ifstop(struct ifnet *, int);
190 static int bcmeth_ifioctl(struct ifnet *, u_long, void *);
191 
192 static int bcmeth_mapcache_create(struct bcmeth_softc *,
193     struct bcmeth_mapcache **, size_t, size_t, size_t);
194 static void bcmeth_mapcache_destroy(struct bcmeth_softc *,
195     struct bcmeth_mapcache *);
196 static bus_dmamap_t bcmeth_mapcache_get(struct bcmeth_softc *,
197     struct bcmeth_mapcache *);
198 static void bcmeth_mapcache_put(struct bcmeth_softc *,
199     struct bcmeth_mapcache *, bus_dmamap_t);
200 
201 static int bcmeth_txq_attach(struct bcmeth_softc *,
202     struct bcmeth_txqueue *, u_int);
203 static void bcmeth_txq_purge(struct bcmeth_softc *,
204     struct bcmeth_txqueue *);
205 static void bcmeth_txq_reset(struct bcmeth_softc *,
206     struct bcmeth_txqueue *);
207 static bool bcmeth_txq_consume(struct bcmeth_softc *,
208     struct bcmeth_txqueue *);
209 static bool bcmeth_txq_produce(struct bcmeth_softc *,
210     struct bcmeth_txqueue *, struct mbuf *m);
211 static bool bcmeth_txq_active_p(struct bcmeth_softc *,
212     struct bcmeth_txqueue *);
213 
214 static int bcmeth_rxq_attach(struct bcmeth_softc *,
215     struct bcmeth_rxqueue *, u_int);
216 static bool bcmeth_rxq_produce(struct bcmeth_softc *,
217     struct bcmeth_rxqueue *);
218 static void bcmeth_rxq_purge(struct bcmeth_softc *,
219     struct bcmeth_rxqueue *, bool);
220 static void bcmeth_rxq_reset(struct bcmeth_softc *,
221     struct bcmeth_rxqueue *);
222 
223 static int bcmeth_intr(void *);
224 #ifdef BCMETH_MPSAFETX
225 static void bcmeth_soft_txintr(struct bcmeth_softc *);
226 #endif
227 static void bcmeth_soft_intr(void *);
228 static void bcmeth_worker(struct work *, void *);
229 
230 static int bcmeth_mediachange(struct ifnet *);
231 static void bcmeth_mediastatus(struct ifnet *, struct ifmediareq *);
232 
233 static inline uint32_t
234 bcmeth_read_4(struct bcmeth_softc *sc, bus_size_t o)
235 {
236 	return bus_space_read_4(sc->sc_bst, sc->sc_bsh, o);
237 }
238 
239 static inline void
240 bcmeth_write_4(struct bcmeth_softc *sc, bus_size_t o, uint32_t v)
241 {
242 	bus_space_write_4(sc->sc_bst, sc->sc_bsh, o, v);
243 }
244 
245 CFATTACH_DECL_NEW(bcmeth_ccb, sizeof(struct bcmeth_softc),
246 	bcmeth_ccb_match, bcmeth_ccb_attach, NULL, NULL);
247 
248 static int
249 bcmeth_ccb_match(device_t parent, cfdata_t cf, void *aux)
250 {
251 	struct bcmccb_attach_args * const ccbaa = aux;
252 	const struct bcm_locators * const loc = &ccbaa->ccbaa_loc;
253 
254 	if (strcmp(cf->cf_name, loc->loc_name))
255 		return 0;
256 
257 #ifdef DIAGNOSTIC
258 	const int port = cf->cf_loc[BCMCCBCF_PORT];
259 #endif
260 	KASSERT(port == BCMCCBCF_PORT_DEFAULT || port == loc->loc_port);
261 
262 	return 1;
263 }
264 
265 static void
266 bcmeth_ccb_attach(device_t parent, device_t self, void *aux)
267 {
268 	struct bcmeth_softc * const sc = device_private(self);
269 	struct ethercom * const ec = &sc->sc_ec;
270 	struct ifnet * const ifp = &ec->ec_if;
271 	struct bcmccb_attach_args * const ccbaa = aux;
272 	const struct bcm_locators * const loc = &ccbaa->ccbaa_loc;
273 	const char * const xname = device_xname(self);
274 	prop_dictionary_t dict = device_properties(self);
275 	int error;
276 
277 	sc->sc_bst = ccbaa->ccbaa_ccb_bst;
278 	sc->sc_dmat = ccbaa->ccbaa_dmat;
279 	bus_space_subregion(sc->sc_bst, ccbaa->ccbaa_ccb_bsh,
280 	    loc->loc_offset, loc->loc_size, &sc->sc_bsh);
281 
282 	/*
283 	 * We need to use the coherent dma tag for the GMAC.
284 	 */
285 	sc->sc_dmat = &bcm53xx_coherent_dma_tag;
286 #if _ARM32_NEED_BUS_DMA_BOUNCE
287 	if (device_cfdata(self)->cf_flags & 2) {
288 		sc->sc_dmat = &bcm53xx_bounce_dma_tag;
289 	}
290 #endif
291 
292 	prop_data_t eaprop = prop_dictionary_get(dict, "mac-address");
293         if (eaprop == NULL) {
294 		uint32_t mac0 = bcmeth_read_4(sc, UNIMAC_MAC_0);
295 		uint32_t mac1 = bcmeth_read_4(sc, UNIMAC_MAC_1);
296 		if ((mac0 == 0 && mac1 == 0) || (mac1 & 1)) {
297 			aprint_error(": mac-address property is missing\n");
298 			return;
299 		}
300 		sc->sc_enaddr[0] = (mac0 >> 0) & 0xff;
301 		sc->sc_enaddr[1] = (mac0 >> 8) & 0xff;
302 		sc->sc_enaddr[2] = (mac0 >> 16) & 0xff;
303 		sc->sc_enaddr[3] = (mac0 >> 24) & 0xff;
304 		sc->sc_enaddr[4] = (mac1 >> 0) & 0xff;
305 		sc->sc_enaddr[5] = (mac1 >> 8) & 0xff;
306 	} else {
307 		KASSERT(prop_object_type(eaprop) == PROP_TYPE_DATA);
308 		KASSERT(prop_data_size(eaprop) == ETHER_ADDR_LEN);
309 		memcpy(sc->sc_enaddr, prop_data_data_nocopy(eaprop),
310 		    ETHER_ADDR_LEN);
311 	}
312 	sc->sc_dev = self;
313 	sc->sc_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SOFTNET);
314 	sc->sc_hwlock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_VM);
315 
316 	bcmeth_write_4(sc, GMAC_INTMASK, 0);	// disable interrupts
317 
318 	aprint_naive("\n");
319 	aprint_normal(": Gigabit Ethernet Controller\n");
320 
321 	error = bcmeth_rxq_attach(sc, &sc->sc_rxq, 0);
322 	if (error) {
323 		aprint_error(": failed to init rxq: %d\n", error);
324 		goto fail_1;
325 	}
326 
327 	error = bcmeth_txq_attach(sc, &sc->sc_txq, 0);
328 	if (error) {
329 		aprint_error(": failed to init txq: %d\n", error);
330 		goto fail_1;
331 	}
332 
333 	error = bcmeth_mapcache_create(sc, &sc->sc_rx_mapcache,
334 	    BCMETH_MAXRXMBUFS, MCLBYTES, BCMETH_NRXSEGS);
335 	if (error) {
336 		aprint_error(": failed to allocate rx dmamaps: %d\n", error);
337 		goto fail_1;
338 	}
339 
340 	error = bcmeth_mapcache_create(sc, &sc->sc_tx_mapcache,
341 	    BCMETH_MAXTXMBUFS, MCLBYTES, BCMETH_NTXSEGS);
342 	if (error) {
343 		aprint_error(": failed to allocate tx dmamaps: %d\n", error);
344 		goto fail_1;
345 	}
346 
347 	error = workqueue_create(&sc->sc_workq, xname, bcmeth_worker, sc,
348 	    (PRI_USER + MAXPRI_USER) / 2, IPL_NET, WQ_MPSAFE|WQ_PERCPU);
349 	if (error) {
350 		aprint_error(": failed to create workqueue: %d\n", error);
351 		goto fail_2;
352 	}
353 
354 	sc->sc_soft_ih = softint_establish(SOFTINT_MPSAFE | SOFTINT_NET,
355 	    bcmeth_soft_intr, sc);
356 
357 	if (sc->sc_ih == NULL) {
358 		aprint_error_dev(self, "failed to establish interrupt %d\n",
359 		     loc->loc_intrs[0]);
360 		goto fail_3;
361 	}
362 
363 	sc->sc_ih = intr_establish(loc->loc_intrs[0], IPL_VM, IST_LEVEL,
364 	    bcmeth_intr, sc);
365 
366 	if (sc->sc_ih == NULL) {
367 		aprint_error_dev(self, "failed to establish interrupt %d\n",
368 		     loc->loc_intrs[0]);
369 		goto fail_4;
370 	} else {
371 		aprint_normal_dev(self, "interrupting on irq %d\n",
372 		     loc->loc_intrs[0]);
373 	}
374 
375 	aprint_normal_dev(sc->sc_dev, "Ethernet address %s\n",
376 	    ether_sprintf(sc->sc_enaddr));
377 
378 	/*
379 	 * Since each port in plugged into the switch/flow-accelerator,
380 	 * we hard code at Gige Full-Duplex with Flow Control enabled.
381 	 */
382 	int ifmedia = IFM_ETHER|IFM_1000_T|IFM_FDX;
383 	//ifmedia |= IFM_FLOW|IFM_ETH_TXPAUSE|IFM_ETH_RXPAUSE;
384 	ifmedia_init(&sc->sc_media, IFM_IMASK, bcmeth_mediachange,
385 	    bcmeth_mediastatus);
386 	ifmedia_add(&sc->sc_media, ifmedia, 0, NULL);
387 	ifmedia_set(&sc->sc_media, ifmedia);
388 
389 	ec->ec_capabilities = ETHERCAP_VLAN_MTU | ETHERCAP_JUMBO_MTU;
390 
391 	strlcpy(ifp->if_xname, xname, IFNAMSIZ);
392 	ifp->if_softc = sc;
393 	ifp->if_baudrate = IF_Mbps(1000);
394 	ifp->if_capabilities = 0;
395 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
396 #ifdef BCMETH_MPSAFE
397 	ifp->if_flags2 = IFF2_MPSAFE;
398 #endif
399 	ifp->if_ioctl = bcmeth_ifioctl;
400 	ifp->if_start = bcmeth_ifstart;
401 	ifp->if_watchdog = bcmeth_ifwatchdog;
402 	ifp->if_init = bcmeth_ifinit;
403 	ifp->if_stop = bcmeth_ifstop;
404 	IFQ_SET_READY(&ifp->if_snd);
405 
406 	bcmeth_ifstop(ifp, true);
407 
408 	/*
409 	 * Attach the interface.
410 	 */
411 	error = if_initialize(ifp);
412 	if (error != 0) {
413 		aprint_error_dev(sc->sc_dev, "if_initialize failed(%d)\n",
414 		    error);
415 		goto fail_5;
416 	}
417 	ether_ifattach(ifp, sc->sc_enaddr);
418 	if_register(ifp);
419 
420 #ifdef BCMETH_COUNTERS
421 	evcnt_attach_dynamic(&sc->sc_ev_intr, EVCNT_TYPE_INTR,
422 	    NULL, xname, "intr");
423 	evcnt_attach_dynamic(&sc->sc_ev_soft_intr, EVCNT_TYPE_INTR,
424 	    NULL, xname, "soft intr");
425 	evcnt_attach_dynamic(&sc->sc_ev_work, EVCNT_TYPE_MISC,
426 	    NULL, xname, "work items");
427 	evcnt_attach_dynamic(&sc->sc_ev_tx_stall, EVCNT_TYPE_MISC,
428 	    NULL, xname, "tx stalls");
429 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_lo, EVCNT_TYPE_MISC,
430 	    NULL, xname, "rx badmagic lo");
431 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_hi, EVCNT_TYPE_MISC,
432 	    NULL, xname, "rx badmagic hi");
433 #endif
434 
435 	return;
436 
437 fail_5:
438 	ifmedia_removeall(&sc->sc_media);
439 fail_4:
440 	intr_disestablish(sc->sc_ih);
441 fail_3:
442 	softint_disestablish(sc->sc_soft_ih);
443 fail_2:
444 	workqueue_destroy(sc->sc_workq);
445 fail_1:
446 	mutex_obj_free(sc->sc_lock);
447 	mutex_obj_free(sc->sc_hwlock);
448 }
449 
450 static int
451 bcmeth_mediachange(struct ifnet *ifp)
452 {
453 	//struct bcmeth_softc * const sc = ifp->if_softc;
454 	return 0;
455 }
456 
457 static void
458 bcmeth_mediastatus(struct ifnet *ifp, struct ifmediareq *ifm)
459 {
460 	//struct bcmeth_softc * const sc = ifp->if_softc;
461 
462 	ifm->ifm_status = IFM_AVALID | IFM_ACTIVE;
463 	ifm->ifm_active = IFM_ETHER | IFM_FDX | IFM_1000_T;
464 }
465 
466 static uint64_t
467 bcmeth_macaddr_create(const uint8_t *enaddr)
468 {
469 	return (enaddr[3] << 0)			// UNIMAC_MAC_0
470 	    |  (enaddr[2] << 8)			// UNIMAC_MAC_0
471 	    |  (enaddr[1] << 16)		// UNIMAC_MAC_0
472 	    |  ((uint64_t)enaddr[0] << 24)	// UNIMAC_MAC_0
473 	    |  ((uint64_t)enaddr[5] << 32)	// UNIMAC_MAC_1
474 	    |  ((uint64_t)enaddr[4] << 40);	// UNIMAC_MAC_1
475 }
476 
477 static int
478 bcmeth_ifinit(struct ifnet *ifp)
479 {
480 	struct bcmeth_softc * const sc = ifp->if_softc;
481 	int error = 0;
482 
483 	sc->sc_maxfrm = max(ifp->if_mtu + 32, MCLBYTES);
484 	if (ifp->if_mtu > ETHERMTU_JUMBO)
485 		return error;
486 
487 	KASSERT(ifp->if_flags & IFF_UP);
488 
489 	/*
490 	 * Stop the interface
491 	 */
492 	bcmeth_ifstop(ifp, 0);
493 
494 	/*
495 	 * Reserve enough space at the front so that we can insert a maxsized
496 	 * link header and a VLAN tag.  Also make sure we have enough room for
497 	 * the rcvsts field as well.
498 	 */
499 	KASSERT(ALIGN(max_linkhdr) == max_linkhdr);
500 	KASSERTMSG(max_linkhdr > sizeof(struct ether_header), "%u > %zu",
501 	    max_linkhdr, sizeof(struct ether_header));
502 	sc->sc_rcvoffset = max_linkhdr + 4 - sizeof(struct ether_header);
503 	if (sc->sc_rcvoffset <= 4)
504 		sc->sc_rcvoffset += 4;
505 	KASSERT((sc->sc_rcvoffset & 3) == 2);
506 	KASSERT(sc->sc_rcvoffset <= __SHIFTOUT(RCVCTL_RCVOFFSET, RCVCTL_RCVOFFSET));
507 	KASSERT(sc->sc_rcvoffset >= 6);
508 
509 	/*
510 	 * If our frame size has changed (or it's our first time through)
511 	 * destroy the existing transmit mapcache.
512 	 */
513 	if (sc->sc_tx_mapcache != NULL
514 	    && sc->sc_maxfrm != sc->sc_tx_mapcache->dmc_maxmapsize) {
515 		bcmeth_mapcache_destroy(sc, sc->sc_tx_mapcache);
516 		sc->sc_tx_mapcache = NULL;
517 	}
518 
519 	if (sc->sc_tx_mapcache == NULL) {
520 		error = bcmeth_mapcache_create(sc, &sc->sc_tx_mapcache,
521 		    BCMETH_MAXTXMBUFS, sc->sc_maxfrm, BCMETH_NTXSEGS);
522 		if (error)
523 			return error;
524 	}
525 
526 	sc->sc_cmdcfg = NO_LENGTH_CHECK | PAUSE_IGNORE
527 	    | __SHIFTIN(ETH_SPEED_1000, ETH_SPEED)
528 	    | RX_ENA | TX_ENA;
529 
530 	if (ifp->if_flags & IFF_PROMISC) {
531 		sc->sc_cmdcfg |= PROMISC_EN;
532 	} else {
533 		sc->sc_cmdcfg &= ~PROMISC_EN;
534 	}
535 
536 	const uint8_t * const lladdr = CLLADDR(ifp->if_sadl);
537 	const uint64_t macstnaddr = bcmeth_macaddr_create(lladdr);
538 
539 	/*
540 	 * We make sure that a received Ethernet packet start on a non-word
541 	 * boundary so that the packet payload will be on a word boundary.
542 	 * So to check the destination address we keep around two words to
543 	 * quickly compare with.
544 	 */
545 #if __ARMEL__
546 	sc->sc_macaddr[0] = lladdr[0] | (lladdr[1] << 8);
547 	sc->sc_macaddr[1] = lladdr[2] | (lladdr[3] << 8)
548 	    | (lladdr[4] << 16) | (lladdr[5] << 24);
549 #else
550 	sc->sc_macaddr[0] = lladdr[1] | (lladdr[0] << 8);
551 	sc->sc_macaddr[1] = lladdr[5] | (lladdr[4] << 8)
552 	    | (lladdr[1] << 16) | (lladdr[2] << 24);
553 #endif
554 
555 	sc->sc_intmask = DESCPROTOERR|DATAERR|DESCERR;
556 
557 	/* 5. Load RCVADDR_LO with new pointer */
558 	bcmeth_rxq_reset(sc, &sc->sc_rxq);
559 
560 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
561 	    __SHIFTIN(sc->sc_rcvoffset, RCVCTL_RCVOFFSET)
562 	    | RCVCTL_PARITY_DIS
563 	    | RCVCTL_OFLOW_CONTINUE
564 	    | __SHIFTIN(3, RCVCTL_BURSTLEN));
565 
566 	/* 6. Load XMTADDR_LO with new pointer */
567 	bcmeth_txq_reset(sc, &sc->sc_txq);
568 
569 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl, XMTCTL_DMA_ACT_INDEX
570 	    | XMTCTL_PARITY_DIS
571 	    | __SHIFTIN(3, XMTCTL_BURSTLEN));
572 
573 	/* 7. Setup other UNIMAC registers */
574 	bcmeth_write_4(sc, UNIMAC_FRAME_LEN, sc->sc_maxfrm);
575 	bcmeth_write_4(sc, UNIMAC_MAC_0, (uint32_t)(macstnaddr >>  0));
576 	bcmeth_write_4(sc, UNIMAC_MAC_1, (uint32_t)(macstnaddr >> 32));
577 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, sc->sc_cmdcfg);
578 
579 	uint32_t devctl = bcmeth_read_4(sc, GMAC_DEVCONTROL);
580 	devctl |= RGMII_LINK_STATUS_SEL | NWAY_AUTO_POLL_EN | TXARB_STRICT_MODE;
581 	devctl &= ~FLOW_CTRL_MODE;
582 	devctl &= ~MIB_RD_RESET_EN;
583 	devctl &= ~RXQ_OVERFLOW_CTRL_SEL;
584 	devctl &= ~CPU_FLOW_CTRL_ON;
585 	bcmeth_write_4(sc, GMAC_DEVCONTROL, devctl);
586 
587 	/* Setup lazy receive (at most 1ms). */
588 	const struct cpu_softc * const cpu = curcpu()->ci_softc;
589 	sc->sc_rcvlazy =  __SHIFTIN(4, INTRCVLAZY_FRAMECOUNT)
590 	     | __SHIFTIN(cpu->cpu_clk.clk_apb / 1000, INTRCVLAZY_TIMEOUT);
591 	bcmeth_write_4(sc, GMAC_INTRCVLAZY, sc->sc_rcvlazy);
592 
593 	/* 11. Enable transmit queues in TQUEUE, and ensure that the transmit scheduling mode is correctly set in TCTRL. */
594 	sc->sc_intmask |= XMTINT_0|XMTUF;
595 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl,
596 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl) | XMTCTL_ENABLE);
597 
598 
599 	/* 12. Enable receive queues in RQUEUE, */
600 	sc->sc_intmask |= RCVINT|RCVDESCUF|RCVFIFOOF;
601 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
602 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl) | RCVCTL_ENABLE);
603 
604 	bcmeth_rxq_produce(sc, &sc->sc_rxq);	/* fill with rx buffers */
605 
606 #if 0
607 	aprint_normal_dev(sc->sc_dev,
608 	    "devctl=%#x ucmdcfg=%#x xmtctl=%#x rcvctl=%#x\n",
609 	    devctl, sc->sc_cmdcfg,
610 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl),
611 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl));
612 #endif
613 
614 	sc->sc_soft_flags = 0;
615 
616 	bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
617 
618 	ifp->if_flags |= IFF_RUNNING;
619 
620 	return error;
621 }
622 
623 static void
624 bcmeth_ifstop(struct ifnet *ifp, int disable)
625 {
626 	struct bcmeth_softc * const sc = ifp->if_softc;
627 	struct bcmeth_txqueue * const txq = &sc->sc_txq;
628 	struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
629 
630 	KASSERT(!cpu_intr_p());
631 
632 	sc->sc_soft_flags = 0;
633 	sc->sc_work_flags = 0;
634 
635 	/* Disable Rx processing */
636 	bcmeth_write_4(sc, rxq->rxq_reg_rcvctl,
637 	    bcmeth_read_4(sc, rxq->rxq_reg_rcvctl) & ~RCVCTL_ENABLE);
638 
639 	/* Disable Tx processing */
640 	bcmeth_write_4(sc, txq->txq_reg_xmtctl,
641 	    bcmeth_read_4(sc, txq->txq_reg_xmtctl) & ~XMTCTL_ENABLE);
642 
643 	/* Disable all interrupts */
644 	bcmeth_write_4(sc, GMAC_INTMASK, 0);
645 
646 	for (;;) {
647 		uint32_t tx0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
648 		uint32_t rx0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
649 		if (__SHIFTOUT(tx0, XMTSTATE) == XMTSTATE_DIS
650 		    && __SHIFTOUT(rx0, RCVSTATE) == RCVSTATE_DIS)
651 			break;
652 		delay(50);
653 	}
654 	/*
655 	 * Now reset the controller.
656 	 *
657 	 * 3. Set SW_RESET bit in UNIMAC_COMMAND_CONFIG register
658 	 * 4. Clear SW_RESET bit in UNIMAC_COMMAND_CONFIG register
659 	 */
660 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, SW_RESET);
661 	bcmeth_write_4(sc, GMAC_INTSTATUS, ~0);
662 	sc->sc_intmask = 0;
663 	ifp->if_flags &= ~IFF_RUNNING;
664 
665 	/*
666 	 * Let's consume any remaining transmitted packets.  And if we are
667 	 * disabling the interface, purge ourselves of any untransmitted
668 	 * packets.  But don't consume any received packets, just drop them.
669 	 * If we aren't disabling the interface, save the mbufs in the
670 	 * receive queue for reuse.
671 	 */
672 	bcmeth_rxq_purge(sc, &sc->sc_rxq, disable);
673 	bcmeth_txq_consume(sc, &sc->sc_txq);
674 	if (disable) {
675 		bcmeth_txq_purge(sc, &sc->sc_txq);
676 		IF_PURGE(&ifp->if_snd);
677 	}
678 
679 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, 0);
680 }
681 
682 static void
683 bcmeth_ifwatchdog(struct ifnet *ifp)
684 {
685 }
686 
687 static int
688 bcmeth_ifioctl(struct ifnet *ifp, u_long cmd, void *data)
689 {
690 	struct bcmeth_softc *sc  = ifp->if_softc;
691 	struct ifreq * const ifr = data;
692 	const int s = splnet();
693 	int error;
694 
695 	switch (cmd) {
696 	case SIOCSIFMEDIA:
697 	case SIOCGIFMEDIA:
698 		error = ifmedia_ioctl(ifp, ifr, &sc->sc_media, cmd);
699 		break;
700 
701 	default:
702 		error = ether_ioctl(ifp, cmd, data);
703 		if (error != ENETRESET)
704 			break;
705 
706 		if (cmd == SIOCADDMULTI || cmd == SIOCDELMULTI) {
707 			error = 0;
708 			break;
709 		}
710 		error = bcmeth_ifinit(ifp);
711 		break;
712 	}
713 
714 	splx(s);
715 	return error;
716 }
717 
718 static void
719 bcmeth_rxq_desc_presync(
720 	struct bcmeth_softc *sc,
721 	struct bcmeth_rxqueue *rxq,
722 	struct gmac_rxdb *rxdb,
723 	size_t count)
724 {
725 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
726 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
727 	    BUS_DMASYNC_PREWRITE);
728 }
729 
730 static void
731 bcmeth_rxq_desc_postsync(
732 	struct bcmeth_softc *sc,
733 	struct bcmeth_rxqueue *rxq,
734 	struct gmac_rxdb *rxdb,
735 	size_t count)
736 {
737 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
738 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
739 	    BUS_DMASYNC_POSTWRITE);
740 }
741 
742 static void
743 bcmeth_txq_desc_presync(
744 	struct bcmeth_softc *sc,
745 	struct bcmeth_txqueue *txq,
746 	struct gmac_txdb *txdb,
747 	size_t count)
748 {
749 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
750 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
751 	    BUS_DMASYNC_PREWRITE);
752 }
753 
754 static void
755 bcmeth_txq_desc_postsync(
756 	struct bcmeth_softc *sc,
757 	struct bcmeth_txqueue *txq,
758 	struct gmac_txdb *txdb,
759 	size_t count)
760 {
761 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
762 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
763 	    BUS_DMASYNC_POSTWRITE);
764 }
765 
766 static bus_dmamap_t
767 bcmeth_mapcache_get(
768 	struct bcmeth_softc *sc,
769 	struct bcmeth_mapcache *dmc)
770 {
771 	KASSERT(dmc->dmc_nmaps > 0);
772 	KASSERT(dmc->dmc_maps[dmc->dmc_nmaps-1] != NULL);
773 	return dmc->dmc_maps[--dmc->dmc_nmaps];
774 }
775 
776 static void
777 bcmeth_mapcache_put(
778 	struct bcmeth_softc *sc,
779 	struct bcmeth_mapcache *dmc,
780 	bus_dmamap_t map)
781 {
782 	KASSERT(map != NULL);
783 	KASSERT(dmc->dmc_nmaps < dmc->dmc_maxmaps);
784 	dmc->dmc_maps[dmc->dmc_nmaps++] = map;
785 }
786 
787 static void
788 bcmeth_mapcache_destroy(
789 	struct bcmeth_softc *sc,
790 	struct bcmeth_mapcache *dmc)
791 {
792 	const size_t dmc_size =
793 	    offsetof(struct bcmeth_mapcache, dmc_maps[dmc->dmc_maxmaps]);
794 
795 	for (u_int i = 0; i < dmc->dmc_maxmaps; i++) {
796 		bus_dmamap_destroy(sc->sc_dmat, dmc->dmc_maps[i]);
797 	}
798 	kmem_intr_free(dmc, dmc_size);
799 }
800 
801 static int
802 bcmeth_mapcache_create(
803 	struct bcmeth_softc *sc,
804 	struct bcmeth_mapcache **dmc_p,
805 	size_t maxmaps,
806 	size_t maxmapsize,
807 	size_t maxseg)
808 {
809 	const size_t dmc_size =
810 	    offsetof(struct bcmeth_mapcache, dmc_maps[maxmaps]);
811 	struct bcmeth_mapcache * const dmc =
812 		kmem_intr_zalloc(dmc_size, KM_NOSLEEP);
813 
814 	dmc->dmc_maxmaps = maxmaps;
815 	dmc->dmc_nmaps = maxmaps;
816 	dmc->dmc_maxmapsize = maxmapsize;
817 	dmc->dmc_maxseg = maxseg;
818 
819 	for (u_int i = 0; i < maxmaps; i++) {
820 		int error = bus_dmamap_create(sc->sc_dmat, dmc->dmc_maxmapsize,
821 		     dmc->dmc_maxseg, dmc->dmc_maxmapsize, 0,
822 		     BUS_DMA_WAITOK|BUS_DMA_ALLOCNOW, &dmc->dmc_maps[i]);
823 		if (error) {
824 			aprint_error_dev(sc->sc_dev,
825 			    "failed to creat dma map cache "
826 			    "entry %u of %zu: %d\n",
827 			    i, maxmaps, error);
828 			while (i-- > 0) {
829 				bus_dmamap_destroy(sc->sc_dmat,
830 				    dmc->dmc_maps[i]);
831 			}
832 			kmem_intr_free(dmc, dmc_size);
833 			return error;
834 		}
835 		KASSERT(dmc->dmc_maps[i] != NULL);
836 	}
837 
838 	*dmc_p = dmc;
839 
840 	return 0;
841 }
842 
843 #if 0
844 static void
845 bcmeth_dmamem_free(
846 	bus_dma_tag_t dmat,
847 	size_t map_size,
848 	bus_dma_segment_t *seg,
849 	bus_dmamap_t map,
850 	void *kvap)
851 {
852 	bus_dmamap_destroy(dmat, map);
853 	bus_dmamem_unmap(dmat, kvap, map_size);
854 	bus_dmamem_free(dmat, seg, 1);
855 }
856 #endif
857 
858 static int
859 bcmeth_dmamem_alloc(
860 	bus_dma_tag_t dmat,
861 	size_t map_size,
862 	bus_dma_segment_t *seg,
863 	bus_dmamap_t *map,
864 	void **kvap)
865 {
866 	int error;
867 	int nseg;
868 
869 	*kvap = NULL;
870 	*map = NULL;
871 
872 	error = bus_dmamem_alloc(dmat, map_size, 2*PAGE_SIZE, 0,
873 	   seg, 1, &nseg, 0);
874 	if (error)
875 		return error;
876 
877 	KASSERT(nseg == 1);
878 
879 	error = bus_dmamem_map(dmat, seg, nseg, map_size, (void **)kvap, 0);
880 	if (error == 0) {
881 		error = bus_dmamap_create(dmat, map_size, 1, map_size, 0, 0,
882 		    map);
883 		if (error == 0) {
884 			error = bus_dmamap_load(dmat, *map, *kvap, map_size,
885 			    NULL, 0);
886 			if (error == 0)
887 				return 0;
888 			bus_dmamap_destroy(dmat, *map);
889 			*map = NULL;
890 		}
891 		bus_dmamem_unmap(dmat, *kvap, map_size);
892 		*kvap = NULL;
893 	}
894 	bus_dmamem_free(dmat, seg, nseg);
895 	return 0;
896 }
897 
898 static struct mbuf *
899 bcmeth_rx_buf_alloc(
900 	struct bcmeth_softc *sc)
901 {
902 	struct mbuf *m = m_gethdr(M_DONTWAIT, MT_DATA);
903 	if (m == NULL) {
904 		printf("%s:%d: %s\n", __func__, __LINE__, "m_gethdr");
905 		return NULL;
906 	}
907 	MCLGET(m, M_DONTWAIT);
908 	if ((m->m_flags & M_EXT) == 0) {
909 		printf("%s:%d: %s\n", __func__, __LINE__, "MCLGET");
910 		m_freem(m);
911 		return NULL;
912 	}
913 	m->m_len = m->m_pkthdr.len = m->m_ext.ext_size;
914 
915 	bus_dmamap_t map = bcmeth_mapcache_get(sc, sc->sc_rx_mapcache);
916 	if (map == NULL) {
917 		printf("%s:%d: %s\n", __func__, __LINE__, "map get");
918 		m_freem(m);
919 		return NULL;
920 	}
921 	M_SETCTX(m, map);
922 	m->m_len = m->m_pkthdr.len = MCLBYTES;
923 	int error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
924 	    BUS_DMA_READ|BUS_DMA_NOWAIT);
925 	if (error) {
926 		aprint_error_dev(sc->sc_dev, "fail to load rx dmamap: %d\n",
927 		    error);
928 		M_SETCTX(m, NULL);
929 		m_freem(m);
930 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
931 		return NULL;
932 	}
933 	KASSERT(map->dm_mapsize == MCLBYTES);
934 #ifdef BCMETH_RCVMAGIC
935 	*mtod(m, uint32_t *) = htole32(BCMETH_RCVMAGIC);
936 	bus_dmamap_sync(sc->sc_dmat, map, 0, sizeof(uint32_t),
937 	    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
938 	bus_dmamap_sync(sc->sc_dmat, map, sizeof(uint32_t),
939 	    map->dm_mapsize - sizeof(uint32_t), BUS_DMASYNC_PREREAD);
940 #else
941 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
942 	    BUS_DMASYNC_PREREAD);
943 #endif
944 
945 	return m;
946 }
947 
948 static void
949 bcmeth_rx_map_unload(
950 	struct bcmeth_softc *sc,
951 	struct mbuf *m)
952 {
953 	KASSERT(m);
954 	for (; m != NULL; m = m->m_next) {
955 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
956 		KASSERT(map);
957 		KASSERT(map->dm_mapsize == MCLBYTES);
958 		bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_len,
959 		    BUS_DMASYNC_POSTREAD);
960 		bus_dmamap_unload(sc->sc_dmat, map);
961 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
962 		M_SETCTX(m, NULL);
963 	}
964 }
965 
966 static bool
967 bcmeth_rxq_produce(
968 	struct bcmeth_softc *sc,
969 	struct bcmeth_rxqueue *rxq)
970 {
971 	struct gmac_rxdb *producer = rxq->rxq_producer;
972 	bool produced = false;
973 
974 	while (rxq->rxq_inuse < rxq->rxq_threshold) {
975 		struct mbuf *m;
976 		IF_DEQUEUE(&sc->sc_rx_bufcache, m);
977 		if (m == NULL) {
978 			m = bcmeth_rx_buf_alloc(sc);
979 			if (m == NULL) {
980 				printf("%s: bcmeth_rx_buf_alloc failed\n", __func__);
981 				break;
982 			}
983 		}
984 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
985 		KASSERT(map);
986 
987 		producer->rxdb_buflen = htole32(MCLBYTES);
988 		producer->rxdb_addrlo = htole32(map->dm_segs[0].ds_addr);
989 		producer->rxdb_flags &= htole32(RXDB_FLAG_ET);
990 		*rxq->rxq_mtail = m;
991 		rxq->rxq_mtail = &m->m_next;
992 		m->m_len = MCLBYTES;
993 		m->m_next = NULL;
994 		rxq->rxq_inuse++;
995 		if (++producer == rxq->rxq_last) {
996 			membar_producer();
997 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
998 			    rxq->rxq_last - rxq->rxq_producer);
999 			producer = rxq->rxq_producer = rxq->rxq_first;
1000 		}
1001 		produced = true;
1002 	}
1003 	if (produced) {
1004 		membar_producer();
1005 		if (producer != rxq->rxq_producer) {
1006 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
1007 			    producer - rxq->rxq_producer);
1008 			rxq->rxq_producer = producer;
1009 		}
1010 		bcmeth_write_4(sc, rxq->rxq_reg_rcvptr,
1011 		    rxq->rxq_descmap->dm_segs[0].ds_addr
1012 		    + ((uintptr_t)producer & RCVPTR));
1013 	}
1014 	return true;
1015 }
1016 
1017 static void
1018 bcmeth_rx_input(
1019 	struct bcmeth_softc *sc,
1020 	struct mbuf *m,
1021 	uint32_t rxdb_flags)
1022 {
1023 	struct ifnet * const ifp = &sc->sc_if;
1024 
1025 	bcmeth_rx_map_unload(sc, m);
1026 
1027 	m_adj(m, sc->sc_rcvoffset);
1028 
1029 	/*
1030 	 * If we are in promiscuous mode and this isn't a multicast, check the
1031 	 * destination address to make sure it matches our own.  If it doesn't,
1032 	 * mark the packet as being received promiscuously.
1033 	 */
1034 	if ((sc->sc_cmdcfg & PROMISC_EN)
1035 	    && (m->m_data[0] & 1) == 0
1036 	    && (*(uint16_t *)&m->m_data[0] != sc->sc_macaddr[0]
1037 		|| *(uint32_t *)&m->m_data[2] != sc->sc_macaddr[1])) {
1038 		m->m_flags |= M_PROMISC;
1039 	}
1040 	m_set_rcvif(m, ifp);
1041 
1042 	ifp->if_ibytes += m->m_pkthdr.len;
1043 
1044 	/*
1045 	 * Let's give it to the network subsystm to deal with.
1046 	 */
1047 #ifdef BCMETH_MPSAFE
1048 	mutex_exit(sc->sc_lock);
1049 	if_input(ifp, m);
1050 	mutex_enter(sc->sc_lock);
1051 #else
1052 	int s = splnet();
1053 	if_input(ifp, m);
1054 	splx(s);
1055 #endif
1056 }
1057 
1058 static bool
1059 bcmeth_rxq_consume(
1060 	struct bcmeth_softc *sc,
1061 	struct bcmeth_rxqueue *rxq,
1062 	size_t atmost)
1063 {
1064 	struct ifnet * const ifp = &sc->sc_if;
1065 	struct gmac_rxdb *consumer = rxq->rxq_consumer;
1066 	size_t rxconsumed = 0;
1067 	bool didconsume = false;
1068 
1069 	while (atmost-- > 0) {
1070 		if (consumer == rxq->rxq_producer) {
1071 			KASSERT(rxq->rxq_inuse == 0);
1072 			break;
1073 		}
1074 
1075 		uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1076 		uint32_t currdscr = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1077 		if (consumer == rxq->rxq_first + currdscr) {
1078 			break;
1079 		}
1080 		bcmeth_rxq_desc_postsync(sc, rxq, consumer, 1);
1081 
1082 		/*
1083 		 * We own this packet again.  Copy the rxsts word from it.
1084 		 */
1085 		rxconsumed++;
1086 		didconsume = true;
1087 		uint32_t rxsts;
1088 		KASSERT(rxq->rxq_mhead != NULL);
1089 		bus_dmamap_t map = M_GETCTX(rxq->rxq_mhead, bus_dmamap_t);
1090 		bus_dmamap_sync(sc->sc_dmat, map, 0, arm_dcache_align,
1091 		    BUS_DMASYNC_POSTREAD);
1092 		memcpy(&rxsts, rxq->rxq_mhead->m_data, 4);
1093 		rxsts = le32toh(rxsts);
1094 #if 0
1095 		KASSERTMSG(rxsts != BCMETH_RCVMAGIC, "currdscr=%u consumer=%zd",
1096 		    currdscr, consumer - rxq->rxq_first);
1097 #endif
1098 
1099 		/*
1100 		 * Get the count of descriptors.  Fetch the correct number
1101 		 * of mbufs.
1102 		 */
1103 #ifdef BCMETH_RCVMAGIC
1104 		size_t desc_count = rxsts != BCMETH_RCVMAGIC ? __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1 : 1;
1105 #else
1106 		size_t desc_count = __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1;
1107 #endif
1108 		struct mbuf *m = rxq->rxq_mhead;
1109 		struct mbuf *m_last = m;
1110 		for (size_t i = 1; i < desc_count; i++) {
1111 			if (++consumer == rxq->rxq_last) {
1112 				consumer = rxq->rxq_first;
1113 			}
1114 			KASSERTMSG(consumer != rxq->rxq_first + currdscr,
1115 			    "i=%zu rxsts=%#x desc_count=%zu currdscr=%u consumer=%zd",
1116 			    i, rxsts, desc_count, currdscr,
1117 			    consumer - rxq->rxq_first);
1118 			m_last = m_last->m_next;
1119 		}
1120 
1121 		/*
1122 		 * Now remove it/them from the list of enqueued mbufs.
1123 		 */
1124 		if ((rxq->rxq_mhead = m_last->m_next) == NULL)
1125 			rxq->rxq_mtail = &rxq->rxq_mhead;
1126 		m_last->m_next = NULL;
1127 
1128 #ifdef BCMETH_RCVMAGIC
1129 		if (rxsts == BCMETH_RCVMAGIC) {
1130 			ifp->if_ierrors++;
1131 			if ((m->m_ext.ext_paddr >> 28) == 8) {
1132 				BCMETH_EVCNT_INCR(sc->sc_ev_rx_badmagic_lo);
1133 			} else {
1134 				BCMETH_EVCNT_INCR( sc->sc_ev_rx_badmagic_hi);
1135 			}
1136 			IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1137 		} else
1138 #endif /* BCMETH_RCVMAGIC */
1139 		if (rxsts & (RXSTS_CRC_ERROR|RXSTS_OVERSIZED|RXSTS_PKT_OVERFLOW)) {
1140 			aprint_error_dev(sc->sc_dev, "[%zu]: count=%zu rxsts=%#x\n",
1141 			    consumer - rxq->rxq_first, desc_count, rxsts);
1142 			/*
1143 			 * We encountered an error, take the mbufs and add them
1144 			 * to the rx bufcache so we can quickly reuse them.
1145 			 */
1146 			ifp->if_ierrors++;
1147 			do {
1148 				struct mbuf *m0 = m->m_next;
1149 				m->m_next = NULL;
1150 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1151 				m = m0;
1152 			} while (m);
1153 		} else {
1154 			uint32_t framelen = __SHIFTOUT(rxsts, RXSTS_FRAMELEN);
1155 			framelen += sc->sc_rcvoffset;
1156 			m->m_pkthdr.len = framelen;
1157 			if (desc_count == 1) {
1158 				KASSERT(framelen <= MCLBYTES);
1159 				m->m_len = framelen;
1160 			} else {
1161 				m_last->m_len = framelen & (MCLBYTES - 1);
1162 			}
1163 
1164 #ifdef BCMETH_MPSAFE
1165 			/*
1166 			 * Wrap at the last entry!
1167 			 */
1168 			if (++consumer == rxq->rxq_last) {
1169 				KASSERT(consumer[-1].rxdb_flags & htole32(RXDB_FLAG_ET));
1170 				rxq->rxq_consumer = rxq->rxq_first;
1171 			} else {
1172 				rxq->rxq_consumer = consumer;
1173 			}
1174 			rxq->rxq_inuse -= rxconsumed;
1175 #endif /* BCMETH_MPSAFE */
1176 
1177 			/*
1178 			 * Receive the packet (which releases our lock)
1179 			 */
1180 			bcmeth_rx_input(sc, m, rxsts);
1181 
1182 #ifdef BCMETH_MPSAFE
1183 			/*
1184 			 * Since we had to give up our lock, we need to
1185 			 * refresh these.
1186 			 */
1187 			consumer = rxq->rxq_consumer;
1188 			rxconsumed = 0;
1189 			continue;
1190 #endif /* BCMETH_MPSAFE */
1191 		}
1192 
1193 		/*
1194 		 * Wrap at the last entry!
1195 		 */
1196 		if (++consumer == rxq->rxq_last) {
1197 			KASSERT(consumer[-1].rxdb_flags & htole32(RXDB_FLAG_ET));
1198 			consumer = rxq->rxq_first;
1199 		}
1200 	}
1201 
1202 	/*
1203 	 * Update queue info.
1204 	 */
1205 	rxq->rxq_consumer = consumer;
1206 	rxq->rxq_inuse -= rxconsumed;
1207 
1208 	/*
1209 	 * Did we consume anything?
1210 	 */
1211 	return didconsume;
1212 }
1213 
1214 static void
1215 bcmeth_rxq_purge(
1216 	struct bcmeth_softc *sc,
1217 	struct bcmeth_rxqueue *rxq,
1218 	bool discard)
1219 {
1220 	struct mbuf *m;
1221 
1222 	if ((m = rxq->rxq_mhead) != NULL) {
1223 		if (discard) {
1224 			bcmeth_rx_map_unload(sc, m);
1225 			m_freem(m);
1226 		} else {
1227 			while (m != NULL) {
1228 				struct mbuf *m0 = m->m_next;
1229 				m->m_next = NULL;
1230 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1231 				m = m0;
1232 			}
1233 		}
1234 
1235 	}
1236 
1237 	rxq->rxq_mhead = NULL;
1238 	rxq->rxq_mtail = &rxq->rxq_mhead;
1239 	rxq->rxq_inuse = 0;
1240 }
1241 
1242 static void
1243 bcmeth_rxq_reset(
1244 	struct bcmeth_softc *sc,
1245 	struct bcmeth_rxqueue *rxq)
1246 {
1247 	/*
1248 	 * sync all the descriptors
1249 	 */
1250 	bcmeth_rxq_desc_postsync(sc, rxq, rxq->rxq_first,
1251 	    rxq->rxq_last - rxq->rxq_first);
1252 
1253 	/*
1254 	 * Make sure we own all descriptors in the ring.
1255 	 */
1256 	struct gmac_rxdb *rxdb;
1257 	for (rxdb = rxq->rxq_first; rxdb < rxq->rxq_last - 1; rxdb++) {
1258 		rxdb->rxdb_flags = htole32(RXDB_FLAG_IC);
1259 	}
1260 
1261 	/*
1262 	 * Last descriptor has the wrap flag.
1263 	 */
1264 	rxdb->rxdb_flags = htole32(RXDB_FLAG_ET|RXDB_FLAG_IC);
1265 
1266 	/*
1267 	 * Reset the producer consumer indexes.
1268 	 */
1269 	rxq->rxq_consumer = rxq->rxq_first;
1270 	rxq->rxq_producer = rxq->rxq_first;
1271 	rxq->rxq_inuse = 0;
1272 	if (rxq->rxq_threshold < BCMETH_MINRXMBUFS)
1273 		rxq->rxq_threshold = BCMETH_MINRXMBUFS;
1274 
1275 	sc->sc_intmask |= RCVINT|RCVFIFOOF|RCVDESCUF;
1276 
1277 	/*
1278 	 * Restart the receiver at the first descriptor
1279 	 */
1280 	bcmeth_write_4(sc, rxq->rxq_reg_rcvaddrlo,
1281 	    rxq->rxq_descmap->dm_segs[0].ds_addr);
1282 }
1283 
1284 static int
1285 bcmeth_rxq_attach(
1286 	struct bcmeth_softc *sc,
1287 	struct bcmeth_rxqueue *rxq,
1288 	u_int qno)
1289 {
1290 	size_t desc_count = BCMETH_RINGSIZE / sizeof(rxq->rxq_first[0]);
1291 	int error;
1292 	void *descs;
1293 
1294 	KASSERT(desc_count == 256 || desc_count == 512);
1295 
1296 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1297 	   &rxq->rxq_descmap_seg, &rxq->rxq_descmap, &descs);
1298 	if (error)
1299 		return error;
1300 
1301 	memset(descs, 0, BCMETH_RINGSIZE);
1302 	rxq->rxq_first = descs;
1303 	rxq->rxq_last = rxq->rxq_first + desc_count;
1304 	rxq->rxq_consumer = descs;
1305 	rxq->rxq_producer = descs;
1306 
1307 	bcmeth_rxq_purge(sc, rxq, true);
1308 	bcmeth_rxq_reset(sc, rxq);
1309 
1310 	rxq->rxq_reg_rcvaddrlo = GMAC_RCVADDR_LOW;
1311 	rxq->rxq_reg_rcvctl = GMAC_RCVCONTROL;
1312 	rxq->rxq_reg_rcvptr = GMAC_RCVPTR;
1313 	rxq->rxq_reg_rcvsts0 = GMAC_RCVSTATUS0;
1314 	rxq->rxq_reg_rcvsts1 = GMAC_RCVSTATUS1;
1315 
1316 	return 0;
1317 }
1318 
1319 static bool
1320 bcmeth_txq_active_p(
1321 	struct bcmeth_softc * const sc,
1322 	struct bcmeth_txqueue *txq)
1323 {
1324 	return !IF_IS_EMPTY(&txq->txq_mbufs);
1325 }
1326 
1327 static bool
1328 bcmeth_txq_fillable_p(
1329 	struct bcmeth_softc * const sc,
1330 	struct bcmeth_txqueue *txq)
1331 {
1332 	return txq->txq_free >= txq->txq_threshold;
1333 }
1334 
1335 static int
1336 bcmeth_txq_attach(
1337 	struct bcmeth_softc *sc,
1338 	struct bcmeth_txqueue *txq,
1339 	u_int qno)
1340 {
1341 	size_t desc_count = BCMETH_RINGSIZE / sizeof(txq->txq_first[0]);
1342 	int error;
1343 	void *descs;
1344 
1345 	KASSERT(desc_count == 256 || desc_count == 512);
1346 
1347 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1348 	   &txq->txq_descmap_seg, &txq->txq_descmap, &descs);
1349 	if (error)
1350 		return error;
1351 
1352 	memset(descs, 0, BCMETH_RINGSIZE);
1353 	txq->txq_first = descs;
1354 	txq->txq_last = txq->txq_first + desc_count;
1355 	txq->txq_consumer = descs;
1356 	txq->txq_producer = descs;
1357 
1358 	IFQ_SET_MAXLEN(&txq->txq_mbufs, BCMETH_MAXTXMBUFS);
1359 
1360 	txq->txq_reg_xmtaddrlo = GMAC_XMTADDR_LOW;
1361 	txq->txq_reg_xmtctl = GMAC_XMTCONTROL;
1362 	txq->txq_reg_xmtptr = GMAC_XMTPTR;
1363 	txq->txq_reg_xmtsts0 = GMAC_XMTSTATUS0;
1364 	txq->txq_reg_xmtsts1 = GMAC_XMTSTATUS1;
1365 
1366 	bcmeth_txq_reset(sc, txq);
1367 
1368 	return 0;
1369 }
1370 
1371 static int
1372 bcmeth_txq_map_load(
1373 	struct bcmeth_softc *sc,
1374 	struct bcmeth_txqueue *txq,
1375 	struct mbuf *m)
1376 {
1377 	bus_dmamap_t map;
1378 	int error;
1379 
1380 	map = M_GETCTX(m, bus_dmamap_t);
1381 	if (map != NULL)
1382 		return 0;
1383 
1384 	map = bcmeth_mapcache_get(sc, sc->sc_tx_mapcache);
1385 	if (map == NULL)
1386 		return ENOMEM;
1387 
1388 	error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
1389 	    BUS_DMA_WRITE | BUS_DMA_NOWAIT);
1390 	if (error)
1391 		return error;
1392 
1393 	bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_pkthdr.len,
1394 	    BUS_DMASYNC_PREWRITE);
1395 	M_SETCTX(m, map);
1396 	return 0;
1397 }
1398 
1399 static void
1400 bcmeth_txq_map_unload(
1401 	struct bcmeth_softc *sc,
1402 	struct bcmeth_txqueue *txq,
1403 	struct mbuf *m)
1404 {
1405 	KASSERT(m);
1406 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1407 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1408 	    BUS_DMASYNC_POSTWRITE);
1409 	bus_dmamap_unload(sc->sc_dmat, map);
1410 	bcmeth_mapcache_put(sc, sc->sc_tx_mapcache, map);
1411 }
1412 
1413 static bool
1414 bcmeth_txq_produce(
1415 	struct bcmeth_softc *sc,
1416 	struct bcmeth_txqueue *txq,
1417 	struct mbuf *m)
1418 {
1419 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1420 
1421 	if (map->dm_nsegs > txq->txq_free)
1422 		return false;
1423 
1424 	/*
1425 	 * TCP Offload flag must be set in the first descriptor.
1426 	 */
1427 	struct gmac_txdb *producer = txq->txq_producer;
1428 	uint32_t first_flags = TXDB_FLAG_SF;
1429 	uint32_t last_flags = TXDB_FLAG_EF;
1430 
1431 	/*
1432 	 * If we've produced enough descriptors without consuming any
1433 	 * we need to ask for an interrupt to reclaim some.
1434 	 */
1435 	txq->txq_lastintr += map->dm_nsegs;
1436 	if (txq->txq_lastintr >= txq->txq_threshold
1437 	    || txq->txq_mbufs.ifq_len + 1 == txq->txq_mbufs.ifq_maxlen) {
1438 		txq->txq_lastintr = 0;
1439 		last_flags |= TXDB_FLAG_IC;
1440 	}
1441 
1442 	KASSERT(producer != txq->txq_last);
1443 
1444 	struct gmac_txdb *start = producer;
1445 	size_t count = map->dm_nsegs;
1446 	producer->txdb_flags |= htole32(first_flags);
1447 	producer->txdb_addrlo = htole32(map->dm_segs[0].ds_addr);
1448 	producer->txdb_buflen = htole32(map->dm_segs[0].ds_len);
1449 	for (u_int i = 1; i < map->dm_nsegs; i++) {
1450 #if 0
1451 		printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1452 		    le32toh(producer->txdb_flags),
1453 		    le32toh(producer->txdb_buflen),
1454 		    le32toh(producer->txdb_addrlo),
1455 		    le32toh(producer->txdb_addrhi));
1456 #endif
1457 		if (__predict_false(++producer == txq->txq_last)) {
1458 			bcmeth_txq_desc_presync(sc, txq, start,
1459 			    txq->txq_last - start);
1460 			count -= txq->txq_last - start;
1461 			producer = txq->txq_first;
1462 			start = txq->txq_first;
1463 		}
1464 		producer->txdb_addrlo = htole32(map->dm_segs[i].ds_addr);
1465 		producer->txdb_buflen = htole32(map->dm_segs[i].ds_len);
1466 	}
1467 	producer->txdb_flags |= htole32(last_flags);
1468 #if 0
1469 	printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1470 	    le32toh(producer->txdb_flags), le32toh(producer->txdb_buflen),
1471 	    le32toh(producer->txdb_addrlo), le32toh(producer->txdb_addrhi));
1472 #endif
1473 	if (count)
1474 		bcmeth_txq_desc_presync(sc, txq, start, count);
1475 
1476 	/*
1477 	 * Reduce free count by the number of segments we consumed.
1478 	 */
1479 	txq->txq_free -= map->dm_nsegs;
1480 	KASSERT(map->dm_nsegs == 1 || txq->txq_producer != producer);
1481 	KASSERT(map->dm_nsegs == 1 || (txq->txq_producer->txdb_flags & htole32(TXDB_FLAG_EF)) == 0);
1482 	KASSERT(producer->txdb_flags & htole32(TXDB_FLAG_EF));
1483 
1484 #if 0
1485 	printf("%s: mbuf %p: produced a %u byte packet in %u segments (%zd..%zd)\n",
1486 	    __func__, m, m->m_pkthdr.len, map->dm_nsegs,
1487 	    txq->txq_producer - txq->txq_first, producer - txq->txq_first);
1488 #endif
1489 
1490 	if (producer + 1 == txq->txq_last)
1491 		txq->txq_producer = txq->txq_first;
1492 	else
1493 		txq->txq_producer = producer + 1;
1494 	IF_ENQUEUE(&txq->txq_mbufs, m);
1495 
1496 	/*
1497 	 * Let the transmitter know there's more to do
1498 	 */
1499 	bcmeth_write_4(sc, txq->txq_reg_xmtptr,
1500 	    txq->txq_descmap->dm_segs[0].ds_addr
1501 	    + ((uintptr_t)txq->txq_producer & XMT_LASTDSCR));
1502 
1503 	return true;
1504 }
1505 
1506 static struct mbuf *
1507 bcmeth_copy_packet(struct mbuf *m)
1508 {
1509 	struct mbuf *mext = NULL;
1510 	size_t misalignment = 0;
1511 	size_t hlen = 0;
1512 
1513 	for (mext = m; mext != NULL; mext = mext->m_next) {
1514 		if (mext->m_flags & M_EXT) {
1515 			misalignment = mtod(mext, vaddr_t) & arm_dcache_align;
1516 			break;
1517 		}
1518 		hlen += m->m_len;
1519 	}
1520 
1521 	struct mbuf *n = m->m_next;
1522 	if (m != mext && hlen + misalignment <= MHLEN && false) {
1523 		KASSERT(m->m_pktdat <= m->m_data && m->m_data <= &m->m_pktdat[MHLEN - m->m_len]);
1524 		size_t oldoff = m->m_data - m->m_pktdat;
1525 		size_t off;
1526 		if (mext == NULL) {
1527 			off = (oldoff + hlen > MHLEN) ? 0 : oldoff;
1528 		} else {
1529 			off = MHLEN - (hlen + misalignment);
1530 		}
1531 		KASSERT(off + hlen + misalignment <= MHLEN);
1532 		if (((oldoff ^ off) & arm_dcache_align) != 0 || off < oldoff) {
1533 			memmove(&m->m_pktdat[off], m->m_data, m->m_len);
1534 			m->m_data = &m->m_pktdat[off];
1535 		}
1536 		m_copydata(n, 0, hlen - m->m_len, &m->m_data[m->m_len]);
1537 		m->m_len = hlen;
1538 		m->m_next = mext;
1539 		while (n != mext) {
1540 			n = m_free(n);
1541 		}
1542 		return m;
1543 	}
1544 
1545 	struct mbuf *m0 = m_gethdr(M_DONTWAIT, m->m_type);
1546 	if (m0 == NULL) {
1547 		return NULL;
1548 	}
1549 	M_COPY_PKTHDR(m0, m);
1550 	MCLAIM(m0, m->m_owner);
1551 	if (m0->m_pkthdr.len > MHLEN) {
1552 		MCLGET(m0, M_DONTWAIT);
1553 		if ((m0->m_flags & M_EXT) == 0) {
1554 			m_freem(m0);
1555 			return NULL;
1556 		}
1557 	}
1558 	m0->m_len = m->m_pkthdr.len;
1559 	m_copydata(m, 0, m0->m_len, mtod(m0, void *));
1560 	m_freem(m);
1561 	return m0;
1562 }
1563 
1564 static bool
1565 bcmeth_txq_enqueue(
1566 	struct bcmeth_softc *sc,
1567 	struct bcmeth_txqueue *txq)
1568 {
1569 	for (;;) {
1570 		if (IF_QFULL(&txq->txq_mbufs))
1571 			return false;
1572 		struct mbuf *m = txq->txq_next;
1573 		if (m == NULL) {
1574 			int s = splnet();
1575 			IF_DEQUEUE(&sc->sc_if.if_snd, m);
1576 			splx(s);
1577 			if (m == NULL)
1578 				return true;
1579 			M_SETCTX(m, NULL);
1580 		} else {
1581 			txq->txq_next = NULL;
1582 		}
1583 		/*
1584 		 * If LINK2 is set and this packet uses multiple mbufs,
1585 		 * consolidate it into a single mbuf.
1586 		 */
1587 		if (m->m_next != NULL && (sc->sc_if.if_flags & IFF_LINK2)) {
1588 			struct mbuf *m0 = bcmeth_copy_packet(m);
1589 			if (m0 == NULL) {
1590 				txq->txq_next = m;
1591 				return true;
1592 			}
1593 			m = m0;
1594 		}
1595 		int error = bcmeth_txq_map_load(sc, txq, m);
1596 		if (error) {
1597 			aprint_error_dev(sc->sc_dev,
1598 			    "discarded packet due to "
1599 			    "dmamap load failure: %d\n", error);
1600 			m_freem(m);
1601 			continue;
1602 		}
1603 		KASSERT(txq->txq_next == NULL);
1604 		if (!bcmeth_txq_produce(sc, txq, m)) {
1605 			txq->txq_next = m;
1606 			return false;
1607 		}
1608 		KASSERT(txq->txq_next == NULL);
1609 	}
1610 }
1611 
1612 static bool
1613 bcmeth_txq_consume(
1614 	struct bcmeth_softc *sc,
1615 	struct bcmeth_txqueue *txq)
1616 {
1617 	struct ifnet * const ifp = &sc->sc_if;
1618 	struct gmac_txdb *consumer = txq->txq_consumer;
1619 	size_t txfree = 0;
1620 
1621 #if 0
1622 	printf("%s: entry: free=%zu\n", __func__, txq->txq_free);
1623 #endif
1624 
1625 	for (;;) {
1626 		if (consumer == txq->txq_producer) {
1627 			txq->txq_consumer = consumer;
1628 			txq->txq_free += txfree;
1629 			txq->txq_lastintr -= min(txq->txq_lastintr, txfree);
1630 #if 0
1631 			printf("%s: empty: freed %zu descriptors going from %zu to %zu\n",
1632 			    __func__, txfree, txq->txq_free - txfree, txq->txq_free);
1633 #endif
1634 			KASSERT(txq->txq_lastintr == 0);
1635 			KASSERT(txq->txq_free == txq->txq_last - txq->txq_first - 1);
1636 			return true;
1637 		}
1638 		bcmeth_txq_desc_postsync(sc, txq, consumer, 1);
1639 		uint32_t s0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
1640 		if (consumer == txq->txq_first + __SHIFTOUT(s0, XMT_CURRDSCR)) {
1641 			txq->txq_consumer = consumer;
1642 			txq->txq_free += txfree;
1643 			txq->txq_lastintr -= min(txq->txq_lastintr, txfree);
1644 #if 0
1645 			printf("%s: freed %zu descriptors\n",
1646 			    __func__, txfree);
1647 #endif
1648 			return bcmeth_txq_fillable_p(sc, txq);
1649 		}
1650 
1651 		/*
1652 		 * If this is the last descriptor in the chain, get the
1653 		 * mbuf, free its dmamap, and free the mbuf chain itself.
1654 		 */
1655 		const uint32_t txdb_flags = le32toh(consumer->txdb_flags);
1656 		if (txdb_flags & TXDB_FLAG_EF) {
1657 			struct mbuf *m;
1658 
1659 			IF_DEQUEUE(&txq->txq_mbufs, m);
1660 			KASSERT(m);
1661 			bcmeth_txq_map_unload(sc, txq, m);
1662 #if 0
1663 			printf("%s: mbuf %p: consumed a %u byte packet\n",
1664 			    __func__, m, m->m_pkthdr.len);
1665 #endif
1666 			bpf_mtap(ifp, m);
1667 			ifp->if_opackets++;
1668 			ifp->if_obytes += m->m_pkthdr.len;
1669 			if (m->m_flags & M_MCAST)
1670 				ifp->if_omcasts++;
1671 			m_freem(m);
1672 		}
1673 
1674 		/*
1675 		 * We own this packet again.  Clear all flags except wrap.
1676 		 */
1677 		txfree++;
1678 
1679 		/*
1680 		 * Wrap at the last entry!
1681 		 */
1682 		if (txdb_flags & TXDB_FLAG_ET) {
1683 			consumer->txdb_flags = htole32(TXDB_FLAG_ET);
1684 			KASSERT(consumer + 1 == txq->txq_last);
1685 			consumer = txq->txq_first;
1686 		} else {
1687 			consumer->txdb_flags = 0;
1688 			consumer++;
1689 			KASSERT(consumer < txq->txq_last);
1690 		}
1691 	}
1692 }
1693 
1694 static void
1695 bcmeth_txq_purge(
1696 	struct bcmeth_softc *sc,
1697 	struct bcmeth_txqueue *txq)
1698 {
1699 	struct mbuf *m;
1700 	KASSERT((bcmeth_read_4(sc, UNIMAC_COMMAND_CONFIG) & TX_ENA) == 0);
1701 
1702 	for (;;) {
1703 		IF_DEQUEUE(&txq->txq_mbufs, m);
1704 		if (m == NULL)
1705 			break;
1706 		bcmeth_txq_map_unload(sc, txq, m);
1707 		m_freem(m);
1708 	}
1709 	if ((m = txq->txq_next) != NULL) {
1710 		txq->txq_next = NULL;
1711 		bcmeth_txq_map_unload(sc, txq, m);
1712 		m_freem(m);
1713 	}
1714 }
1715 
1716 static void
1717 bcmeth_txq_reset(
1718 	struct bcmeth_softc *sc,
1719 	struct bcmeth_txqueue *txq)
1720 {
1721 	/*
1722 	 * sync all the descriptors
1723 	 */
1724 	bcmeth_txq_desc_postsync(sc, txq, txq->txq_first,
1725 	    txq->txq_last - txq->txq_first);
1726 
1727 	/*
1728 	 * Make sure we own all descriptors in the ring.
1729 	 */
1730 	struct gmac_txdb *txdb;
1731 	for (txdb = txq->txq_first; txdb < txq->txq_last - 1; txdb++) {
1732 		txdb->txdb_flags = 0;
1733 	}
1734 
1735 	/*
1736 	 * Last descriptor has the wrap flag.
1737 	 */
1738 	txdb->txdb_flags = htole32(TXDB_FLAG_ET);
1739 
1740 	/*
1741 	 * Reset the producer consumer indexes.
1742 	 */
1743 	txq->txq_consumer = txq->txq_first;
1744 	txq->txq_producer = txq->txq_first;
1745 	txq->txq_free = txq->txq_last - txq->txq_first - 1;
1746 	txq->txq_threshold = txq->txq_free / 2;
1747 	txq->txq_lastintr = 0;
1748 
1749 	/*
1750 	 * What do we want to get interrupted on?
1751 	 */
1752 	sc->sc_intmask |= XMTINT_0 | XMTUF;
1753 
1754 	/*
1755 	 * Restart the transmiter at the first descriptor
1756 	 */
1757 	bcmeth_write_4(sc, txq->txq_reg_xmtaddrlo,
1758 	    txq->txq_descmap->dm_segs->ds_addr);
1759 }
1760 
1761 static void
1762 bcmeth_ifstart(struct ifnet *ifp)
1763 {
1764 	struct bcmeth_softc * const sc = ifp->if_softc;
1765 
1766 	if (__predict_false((ifp->if_flags & IFF_RUNNING) == 0)) {
1767 		return;
1768 	}
1769 
1770 #ifdef BCMETH_MPSAFETX
1771 	if (cpu_intr_p()) {
1772 #endif
1773 		atomic_or_uint(&sc->sc_soft_flags, SOFT_TXINTR);
1774 		softint_schedule(sc->sc_soft_ih);
1775 #ifdef BCMETH_MPSAFETX
1776 	} else {
1777 		/*
1778 		 * Either we are in a softintr thread already or some other
1779 		 * thread so just borrow it to do the send and save ourselves
1780 		 * the overhead of a fast soft int.
1781 		 */
1782 		bcmeth_soft_txintr(sc);
1783 	}
1784 #endif
1785 }
1786 
1787 int
1788 bcmeth_intr(void *arg)
1789 {
1790 	struct bcmeth_softc * const sc = arg;
1791 	uint32_t soft_flags = 0;
1792 	uint32_t work_flags = 0;
1793 	int rv = 0;
1794 
1795 	mutex_enter(sc->sc_hwlock);
1796 
1797 	uint32_t intmask = sc->sc_intmask;
1798 	BCMETH_EVCNT_INCR(sc->sc_ev_intr);
1799 
1800 	for (;;) {
1801 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
1802 		intstatus &= intmask;
1803 		bcmeth_write_4(sc, GMAC_INTSTATUS, intstatus);	/* write 1 to clear */
1804 		if (intstatus == 0) {
1805 			break;
1806 		}
1807 #if 0
1808 		aprint_normal_dev(sc->sc_dev, "%s: intstatus=%#x intmask=%#x\n",
1809 		    __func__, intstatus, bcmeth_read_4(sc, GMAC_INTMASK));
1810 #endif
1811 		if (intstatus & RCVINT) {
1812 			struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
1813 			intmask &= ~RCVINT;
1814 
1815 			uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1816 			uint32_t descs = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1817 			if (descs < rxq->rxq_consumer - rxq->rxq_first) {
1818 				/*
1819 				 * We wrapped at the end so count how far
1820 				 * we are from the end.
1821 				 */
1822 				descs += rxq->rxq_last - rxq->rxq_consumer;
1823 			} else {
1824 				descs -= rxq->rxq_consumer - rxq->rxq_first;
1825 			}
1826 			/*
1827 			 * If we "timedout" we can't be hogging so use
1828 			 * softints.  If we exceeded then we might hogging
1829 			 * so let the workqueue deal with them.
1830 			 */
1831 			const uint32_t framecount = __SHIFTOUT(sc->sc_rcvlazy, INTRCVLAZY_FRAMECOUNT);
1832 			if (descs < framecount
1833 			    || (curcpu()->ci_curlwp->l_flag & LW_IDLE)) {
1834 				soft_flags |= SOFT_RXINTR;
1835 			} else {
1836 				work_flags |= WORK_RXINTR;
1837 			}
1838 		}
1839 
1840 		if (intstatus & XMTINT_0) {
1841 			intmask &= ~XMTINT_0;
1842 			soft_flags |= SOFT_TXINTR;
1843 		}
1844 
1845 		if (intstatus & RCVDESCUF) {
1846 			intmask &= ~RCVDESCUF;
1847 			work_flags |= WORK_RXUNDERFLOW;
1848 		}
1849 
1850 		intstatus &= intmask;
1851 		if (intstatus) {
1852 			aprint_error_dev(sc->sc_dev,
1853 			    "intr: intstatus=%#x\n", intstatus);
1854 			aprint_error_dev(sc->sc_dev,
1855 			    "rcvbase=%p/%#lx rcvptr=%#x rcvsts=%#x/%#x\n",
1856 			    sc->sc_rxq.rxq_first,
1857 			    sc->sc_rxq.rxq_descmap->dm_segs[0].ds_addr,
1858 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvptr),
1859 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts0),
1860 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts1));
1861 			aprint_error_dev(sc->sc_dev,
1862 			    "xmtbase=%p/%#lx xmtptr=%#x xmtsts=%#x/%#x\n",
1863 			    sc->sc_txq.txq_first,
1864 			    sc->sc_txq.txq_descmap->dm_segs[0].ds_addr,
1865 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtptr),
1866 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts0),
1867 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts1));
1868 			intmask &= ~intstatus;
1869 			work_flags |= WORK_REINIT;
1870 			break;
1871 		}
1872 	}
1873 
1874 	if (intmask != sc->sc_intmask) {
1875 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1876 	}
1877 
1878 	if (work_flags) {
1879 		if (sc->sc_work_flags == 0) {
1880 			workqueue_enqueue(sc->sc_workq, &sc->sc_work, NULL);
1881 		}
1882 		atomic_or_32(&sc->sc_work_flags, work_flags);
1883 		rv = 1;
1884 	}
1885 
1886 	if (soft_flags) {
1887 		if (sc->sc_soft_flags == 0) {
1888 			softint_schedule(sc->sc_soft_ih);
1889 		}
1890 		atomic_or_32(&sc->sc_soft_flags, soft_flags);
1891 		rv = 1;
1892 	}
1893 
1894 	mutex_exit(sc->sc_hwlock);
1895 
1896 	return rv;
1897 }
1898 
1899 #ifdef BCMETH_MPSAFETX
1900 void
1901 bcmeth_soft_txintr(struct bcmeth_softc *sc)
1902 {
1903 	mutex_enter(sc->sc_lock);
1904 	/*
1905 	 * Let's do what we came here for.  Consume transmitted
1906 	 * packets off the the transmit ring.
1907 	 */
1908 	if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1909 	    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1910 		BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1911 		sc->sc_if.if_flags |= IFF_OACTIVE;
1912 	} else {
1913 		sc->sc_if.if_flags &= ~IFF_OACTIVE;
1914 	}
1915 	if (sc->sc_if.if_flags & IFF_RUNNING) {
1916 		mutex_spin_enter(sc->sc_hwlock);
1917 		sc->sc_intmask |= XMTINT_0;
1918 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1919 		mutex_spin_exit(sc->sc_hwlock);
1920 	}
1921 	mutex_exit(sc->sc_lock);
1922 }
1923 #endif /* BCMETH_MPSAFETX */
1924 
1925 void
1926 bcmeth_soft_intr(void *arg)
1927 {
1928 	struct bcmeth_softc * const sc = arg;
1929 	struct ifnet * const ifp = &sc->sc_if;
1930 	uint32_t intmask = 0;
1931 
1932 	mutex_enter(sc->sc_lock);
1933 
1934 	u_int soft_flags = atomic_swap_uint(&sc->sc_soft_flags, 0);
1935 
1936 	BCMETH_EVCNT_INCR(sc->sc_ev_soft_intr);
1937 
1938 	if ((soft_flags & SOFT_TXINTR)
1939 	    || bcmeth_txq_active_p(sc, &sc->sc_txq)) {
1940 		/*
1941 		 * Let's do what we came here for.  Consume transmitted
1942 		 * packets off the the transmit ring.
1943 		 */
1944 		if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1945 		    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1946 			BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1947 			ifp->if_flags |= IFF_OACTIVE;
1948 		} else {
1949 			ifp->if_flags &= ~IFF_OACTIVE;
1950 		}
1951 		intmask |= XMTINT_0;
1952 	}
1953 
1954 	if (soft_flags & SOFT_RXINTR) {
1955 		/*
1956 		 * Let's consume
1957 		 */
1958 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
1959 		    sc->sc_rxq.rxq_threshold / 4)) {
1960 			/*
1961 			 * We've consumed a quarter of the ring and still have
1962 			 * more to do.  Refill the ring.
1963 			 */
1964 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
1965 		}
1966 		intmask |= RCVINT;
1967 	}
1968 
1969 	if (ifp->if_flags & IFF_RUNNING) {
1970 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
1971 		mutex_spin_enter(sc->sc_hwlock);
1972 		sc->sc_intmask |= intmask;
1973 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1974 		mutex_spin_exit(sc->sc_hwlock);
1975 	}
1976 
1977 	mutex_exit(sc->sc_lock);
1978 }
1979 
1980 void
1981 bcmeth_worker(struct work *wk, void *arg)
1982 {
1983 	struct bcmeth_softc * const sc = arg;
1984 	struct ifnet * const ifp = &sc->sc_if;
1985 	uint32_t intmask = 0;
1986 
1987 	mutex_enter(sc->sc_lock);
1988 
1989 	BCMETH_EVCNT_INCR(sc->sc_ev_work);
1990 
1991 	uint32_t work_flags = atomic_swap_32(&sc->sc_work_flags, 0);
1992 	if (work_flags & WORK_REINIT) {
1993 		int s = splnet();
1994 		sc->sc_soft_flags = 0;
1995 		bcmeth_ifinit(ifp);
1996 		splx(s);
1997 		work_flags &= ~WORK_RXUNDERFLOW;
1998 	}
1999 
2000 	if (work_flags & WORK_RXUNDERFLOW) {
2001 		struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
2002 		size_t threshold = 5 * rxq->rxq_threshold / 4;
2003 		if (threshold >= rxq->rxq_last - rxq->rxq_first) {
2004 			threshold = rxq->rxq_last - rxq->rxq_first - 1;
2005 		} else {
2006 			intmask |= RCVDESCUF;
2007 		}
2008 		aprint_normal_dev(sc->sc_dev,
2009 		    "increasing receive buffers from %zu to %zu\n",
2010 		    rxq->rxq_threshold, threshold);
2011 		rxq->rxq_threshold = threshold;
2012 	}
2013 
2014 	if (work_flags & WORK_RXINTR) {
2015 		/*
2016 		 * Let's consume
2017 		 */
2018 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
2019 		    sc->sc_rxq.rxq_threshold / 4)) {
2020 			/*
2021 			 * We've consumed a quarter of the ring and still have
2022 			 * more to do.  Refill the ring.
2023 			 */
2024 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
2025 		}
2026 		intmask |= RCVINT;
2027 	}
2028 
2029 	if (ifp->if_flags & IFF_RUNNING) {
2030 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
2031 #if 0
2032 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
2033 		if (intstatus & RCVINT) {
2034 			bcmeth_write_4(sc, GMAC_INTSTATUS, RCVINT);
2035 			work_flags |= WORK_RXINTR;
2036 			continue;
2037 		}
2038 #endif
2039 		mutex_spin_enter(sc->sc_hwlock);
2040 		sc->sc_intmask |= intmask;
2041 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
2042 		mutex_spin_exit(sc->sc_hwlock);
2043 	}
2044 
2045 	mutex_exit(sc->sc_lock);
2046 }
2047