xref: /netbsd-src/sys/arch/arm/broadcom/bcm53xx_eth.c (revision 79604517018af1a68810e40d59b9622b26206b51)
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 1.45 2024/12/04 21:18:34 andvar 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 	const int port __diagused = cf->cf_loc[BCMCCBCF_PORT];
256 	KASSERT(port == BCMCCBCF_PORT_DEFAULT || port == loc->loc_port);
257 
258 	return 1;
259 }
260 
261 static void
262 bcmeth_ccb_attach(device_t parent, device_t self, void *aux)
263 {
264 	struct bcmeth_softc * const sc = device_private(self);
265 	struct ethercom * const ec = &sc->sc_ec;
266 	struct ifnet * const ifp = &ec->ec_if;
267 	struct bcmccb_attach_args * const ccbaa = aux;
268 	const struct bcm_locators * const loc = &ccbaa->ccbaa_loc;
269 	const char * const xname = device_xname(self);
270 	prop_dictionary_t dict = device_properties(self);
271 	int error;
272 
273 	sc->sc_bst = ccbaa->ccbaa_ccb_bst;
274 	sc->sc_dmat = ccbaa->ccbaa_dmat;
275 	bus_space_subregion(sc->sc_bst, ccbaa->ccbaa_ccb_bsh,
276 	    loc->loc_offset, loc->loc_size, &sc->sc_bsh);
277 
278 	/*
279 	 * We need to use the coherent dma tag for the GMAC.
280 	 */
281 	sc->sc_dmat = &bcm53xx_coherent_dma_tag;
282 #if _ARM32_NEED_BUS_DMA_BOUNCE
283 	if (device_cfdata(self)->cf_flags & 2) {
284 		sc->sc_dmat = &bcm53xx_bounce_dma_tag;
285 	}
286 #endif
287 
288 	prop_data_t eaprop = prop_dictionary_get(dict, "mac-address");
289 	if (eaprop == NULL) {
290 		uint32_t mac0 = bcmeth_read_4(sc, UNIMAC_MAC_0);
291 		uint32_t mac1 = bcmeth_read_4(sc, UNIMAC_MAC_1);
292 		if ((mac0 == 0 && mac1 == 0) || (mac1 & 1)) {
293 			aprint_error(": mac-address property is missing\n");
294 			return;
295 		}
296 		sc->sc_enaddr[0] = (mac0 >> 0) & 0xff;
297 		sc->sc_enaddr[1] = (mac0 >> 8) & 0xff;
298 		sc->sc_enaddr[2] = (mac0 >> 16) & 0xff;
299 		sc->sc_enaddr[3] = (mac0 >> 24) & 0xff;
300 		sc->sc_enaddr[4] = (mac1 >> 0) & 0xff;
301 		sc->sc_enaddr[5] = (mac1 >> 8) & 0xff;
302 	} else {
303 		KASSERT(prop_object_type(eaprop) == PROP_TYPE_DATA);
304 		KASSERT(prop_data_size(eaprop) == ETHER_ADDR_LEN);
305 		memcpy(sc->sc_enaddr, prop_data_data_nocopy(eaprop),
306 		    ETHER_ADDR_LEN);
307 	}
308 	sc->sc_dev = self;
309 	sc->sc_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SOFTNET);
310 	sc->sc_hwlock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_VM);
311 
312 	bcmeth_write_4(sc, GMAC_INTMASK, 0);	// disable interrupts
313 
314 	aprint_naive("\n");
315 	aprint_normal(": Gigabit Ethernet Controller\n");
316 
317 	error = bcmeth_rxq_attach(sc, &sc->sc_rxq, 0);
318 	if (error) {
319 		aprint_error(": failed to init rxq: %d\n", error);
320 		goto fail_1;
321 	}
322 
323 	error = bcmeth_txq_attach(sc, &sc->sc_txq, 0);
324 	if (error) {
325 		aprint_error(": failed to init txq: %d\n", error);
326 		goto fail_1;
327 	}
328 
329 	error = bcmeth_mapcache_create(sc, &sc->sc_rx_mapcache,
330 	    BCMETH_MAXRXMBUFS, MCLBYTES, BCMETH_NRXSEGS);
331 	if (error) {
332 		aprint_error(": failed to allocate rx dmamaps: %d\n", error);
333 		goto fail_1;
334 	}
335 
336 	error = bcmeth_mapcache_create(sc, &sc->sc_tx_mapcache,
337 	    BCMETH_MAXTXMBUFS, MCLBYTES, BCMETH_NTXSEGS);
338 	if (error) {
339 		aprint_error(": failed to allocate tx dmamaps: %d\n", error);
340 		goto fail_1;
341 	}
342 
343 	error = workqueue_create(&sc->sc_workq, xname, bcmeth_worker, sc,
344 	    (PRI_USER + MAXPRI_USER) / 2, IPL_NET, WQ_MPSAFE|WQ_PERCPU);
345 	if (error) {
346 		aprint_error(": failed to create workqueue: %d\n", error);
347 		goto fail_1;
348 	}
349 
350 	sc->sc_soft_ih = softint_establish(SOFTINT_MPSAFE | SOFTINT_NET,
351 	    bcmeth_soft_intr, sc);
352 
353 	if (sc->sc_soft_ih == NULL) {
354 		aprint_error_dev(self, "failed to establish soft interrupt\n");
355 		goto fail_2;
356 	}
357 
358 	sc->sc_ih = intr_establish(loc->loc_intrs[0], IPL_VM, IST_LEVEL,
359 	    bcmeth_intr, sc);
360 
361 	if (sc->sc_ih == NULL) {
362 		aprint_error_dev(self, "failed to establish interrupt %d\n",
363 		     loc->loc_intrs[0]);
364 		goto fail_3;
365 	} else {
366 		aprint_normal_dev(self, "interrupting on irq %d\n",
367 		     loc->loc_intrs[0]);
368 	}
369 
370 	aprint_normal_dev(sc->sc_dev, "Ethernet address %s\n",
371 	    ether_sprintf(sc->sc_enaddr));
372 
373 	/*
374 	 * Since each port in plugged into the switch/flow-accelerator,
375 	 * we hard code at Gige Full-Duplex with Flow Control enabled.
376 	 */
377 	int ifmedia = IFM_ETHER | IFM_1000_T | IFM_FDX;
378 	//ifmedia |= IFM_FLOW | IFM_ETH_TXPAUSE | IFM_ETH_RXPAUSE;
379 	ec->ec_ifmedia = &sc->sc_media;
380 	ifmedia_init(&sc->sc_media, IFM_IMASK, bcmeth_mediachange,
381 	    bcmeth_mediastatus);
382 	ifmedia_add(&sc->sc_media, ifmedia, 0, NULL);
383 	ifmedia_set(&sc->sc_media, ifmedia);
384 
385 	ec->ec_capabilities = ETHERCAP_VLAN_MTU | ETHERCAP_JUMBO_MTU;
386 
387 	strlcpy(ifp->if_xname, xname, IFNAMSIZ);
388 	ifp->if_softc = sc;
389 	ifp->if_baudrate = IF_Mbps(1000);
390 	ifp->if_capabilities = 0;
391 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
392 #ifdef BCMETH_MPSAFE
393 	ifp->if_flags2 = IFF2_MPSAFE;
394 #endif
395 	ifp->if_ioctl = bcmeth_ifioctl;
396 	ifp->if_start = bcmeth_ifstart;
397 	ifp->if_watchdog = bcmeth_ifwatchdog;
398 	ifp->if_init = bcmeth_ifinit;
399 	ifp->if_stop = bcmeth_ifstop;
400 	IFQ_SET_READY(&ifp->if_snd);
401 
402 	bcmeth_ifstop(ifp, true);
403 
404 	/*
405 	 * Attach the interface.
406 	 */
407 	if_initialize(ifp);
408 	ether_ifattach(ifp, sc->sc_enaddr);
409 	if_register(ifp);
410 
411 #ifdef BCMETH_COUNTERS
412 	evcnt_attach_dynamic(&sc->sc_ev_intr, EVCNT_TYPE_INTR,
413 	    NULL, xname, "intr");
414 	evcnt_attach_dynamic(&sc->sc_ev_soft_intr, EVCNT_TYPE_INTR,
415 	    NULL, xname, "soft intr");
416 	evcnt_attach_dynamic(&sc->sc_ev_work, EVCNT_TYPE_MISC,
417 	    NULL, xname, "work items");
418 	evcnt_attach_dynamic(&sc->sc_ev_tx_stall, EVCNT_TYPE_MISC,
419 	    NULL, xname, "tx stalls");
420 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_lo, EVCNT_TYPE_MISC,
421 	    NULL, xname, "rx badmagic lo");
422 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_hi, EVCNT_TYPE_MISC,
423 	    NULL, xname, "rx badmagic hi");
424 #endif
425 
426 	return;
427 
428 fail_3:
429 	softint_disestablish(sc->sc_soft_ih);
430 fail_2:
431 	workqueue_destroy(sc->sc_workq);
432 fail_1:
433 	mutex_obj_free(sc->sc_lock);
434 	mutex_obj_free(sc->sc_hwlock);
435 }
436 
437 static int
438 bcmeth_mediachange(struct ifnet *ifp)
439 {
440 	//struct bcmeth_softc * const sc = ifp->if_softc;
441 	return 0;
442 }
443 
444 static void
445 bcmeth_mediastatus(struct ifnet *ifp, struct ifmediareq *ifm)
446 {
447 	//struct bcmeth_softc * const sc = ifp->if_softc;
448 
449 	ifm->ifm_status = IFM_AVALID | IFM_ACTIVE;
450 	ifm->ifm_active = IFM_ETHER | IFM_FDX | IFM_1000_T;
451 }
452 
453 static uint64_t
454 bcmeth_macaddr_create(const uint8_t *enaddr)
455 {
456 	return (enaddr[3] << 0)			// UNIMAC_MAC_0
457 	    |  (enaddr[2] << 8)			// UNIMAC_MAC_0
458 	    |  (enaddr[1] << 16)		// UNIMAC_MAC_0
459 	    |  ((uint64_t)enaddr[0] << 24)	// UNIMAC_MAC_0
460 	    |  ((uint64_t)enaddr[5] << 32)	// UNIMAC_MAC_1
461 	    |  ((uint64_t)enaddr[4] << 40);	// UNIMAC_MAC_1
462 }
463 
464 static int
465 bcmeth_ifinit(struct ifnet *ifp)
466 {
467 	struct bcmeth_softc * const sc = ifp->if_softc;
468 	int error = 0;
469 
470 	sc->sc_maxfrm = uimax(ifp->if_mtu + 32, MCLBYTES);
471 	if (ifp->if_mtu > ETHERMTU_JUMBO)
472 		return error;
473 
474 	KASSERT(ifp->if_flags & IFF_UP);
475 
476 	/*
477 	 * Stop the interface
478 	 */
479 	bcmeth_ifstop(ifp, 0);
480 
481 	/*
482 	 * Reserve enough space at the front so that we can insert a maxsized
483 	 * link header and a VLAN tag.  Also make sure we have enough room for
484 	 * the rcvsts field as well.
485 	 */
486 	KASSERT(ALIGN(max_linkhdr) == max_linkhdr);
487 	KASSERTMSG(max_linkhdr > sizeof(struct ether_header), "%u > %zu",
488 	    max_linkhdr, sizeof(struct ether_header));
489 	sc->sc_rcvoffset = max_linkhdr + 4 - sizeof(struct ether_header);
490 	if (sc->sc_rcvoffset <= 4)
491 		sc->sc_rcvoffset += 4;
492 	KASSERT((sc->sc_rcvoffset & 3) == 2);
493 	KASSERT(sc->sc_rcvoffset <= __SHIFTOUT(RCVCTL_RCVOFFSET, RCVCTL_RCVOFFSET));
494 	KASSERT(sc->sc_rcvoffset >= 6);
495 
496 	/*
497 	 * If our frame size has changed (or it's our first time through)
498 	 * destroy the existing transmit mapcache.
499 	 */
500 	if (sc->sc_tx_mapcache != NULL
501 	    && sc->sc_maxfrm != sc->sc_tx_mapcache->dmc_maxmapsize) {
502 		bcmeth_mapcache_destroy(sc, sc->sc_tx_mapcache);
503 		sc->sc_tx_mapcache = NULL;
504 	}
505 
506 	if (sc->sc_tx_mapcache == NULL) {
507 		error = bcmeth_mapcache_create(sc, &sc->sc_tx_mapcache,
508 		    BCMETH_MAXTXMBUFS, sc->sc_maxfrm, BCMETH_NTXSEGS);
509 		if (error)
510 			return error;
511 	}
512 
513 	sc->sc_cmdcfg = NO_LENGTH_CHECK | PAUSE_IGNORE
514 	    | __SHIFTIN(ETH_SPEED_1000, ETH_SPEED)
515 	    | RX_ENA | TX_ENA;
516 
517 	if (ifp->if_flags & IFF_PROMISC) {
518 		sc->sc_cmdcfg |= PROMISC_EN;
519 	} else {
520 		sc->sc_cmdcfg &= ~PROMISC_EN;
521 	}
522 
523 	const uint8_t * const lladdr = CLLADDR(ifp->if_sadl);
524 	const uint64_t macstnaddr = bcmeth_macaddr_create(lladdr);
525 
526 	/*
527 	 * We make sure that a received Ethernet packet start on a non-word
528 	 * boundary so that the packet payload will be on a word boundary.
529 	 * So to check the destination address we keep around two words to
530 	 * quickly compare with.
531 	 */
532 #if __ARMEL__
533 	sc->sc_macaddr[0] = lladdr[0] | (lladdr[1] << 8);
534 	sc->sc_macaddr[1] = lladdr[2] | (lladdr[3] << 8)
535 	    | (lladdr[4] << 16) | (lladdr[5] << 24);
536 #else
537 	sc->sc_macaddr[0] = lladdr[1] | (lladdr[0] << 8);
538 	sc->sc_macaddr[1] = lladdr[5] | (lladdr[4] << 8)
539 	    | (lladdr[1] << 16) | (lladdr[2] << 24);
540 #endif
541 
542 	sc->sc_intmask = DESCPROTOERR | DATAERR | DESCERR;
543 
544 	/* 5. Load RCVADDR_LO with new pointer */
545 	bcmeth_rxq_reset(sc, &sc->sc_rxq);
546 
547 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
548 	    __SHIFTIN(sc->sc_rcvoffset, RCVCTL_RCVOFFSET)
549 	    | RCVCTL_PARITY_DIS
550 	    | RCVCTL_OFLOW_CONTINUE
551 	    | __SHIFTIN(3, RCVCTL_BURSTLEN));
552 
553 	/* 6. Load XMTADDR_LO with new pointer */
554 	bcmeth_txq_reset(sc, &sc->sc_txq);
555 
556 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl, XMTCTL_DMA_ACT_INDEX
557 	    | XMTCTL_PARITY_DIS
558 	    | __SHIFTIN(3, XMTCTL_BURSTLEN));
559 
560 	/* 7. Setup other UNIMAC registers */
561 	bcmeth_write_4(sc, UNIMAC_FRAME_LEN, sc->sc_maxfrm);
562 	bcmeth_write_4(sc, UNIMAC_MAC_0, (uint32_t)(macstnaddr >>  0));
563 	bcmeth_write_4(sc, UNIMAC_MAC_1, (uint32_t)(macstnaddr >> 32));
564 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, sc->sc_cmdcfg);
565 
566 	uint32_t devctl = bcmeth_read_4(sc, GMAC_DEVCONTROL);
567 	devctl |= RGMII_LINK_STATUS_SEL | NWAY_AUTO_POLL_EN | TXARB_STRICT_MODE;
568 	devctl &= ~FLOW_CTRL_MODE;
569 	devctl &= ~MIB_RD_RESET_EN;
570 	devctl &= ~RXQ_OVERFLOW_CTRL_SEL;
571 	devctl &= ~CPU_FLOW_CTRL_ON;
572 	bcmeth_write_4(sc, GMAC_DEVCONTROL, devctl);
573 
574 	/* Setup lazy receive (at most 1ms). */
575 	const struct cpu_softc * const cpu = curcpu()->ci_softc;
576 	sc->sc_rcvlazy =  __SHIFTIN(4, INTRCVLAZY_FRAMECOUNT)
577 	     | __SHIFTIN(cpu->cpu_clk.clk_apb / 1000, INTRCVLAZY_TIMEOUT);
578 	bcmeth_write_4(sc, GMAC_INTRCVLAZY, sc->sc_rcvlazy);
579 
580 	/* 11. Enable transmit queues in TQUEUE, and ensure that the transmit scheduling mode is correctly set in TCTRL. */
581 	sc->sc_intmask |= XMTINT_0 | XMTUF;
582 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl,
583 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl) | XMTCTL_ENABLE);
584 
585 
586 	/* 12. Enable receive queues in RQUEUE, */
587 	sc->sc_intmask |= RCVINT | RCVDESCUF | RCVFIFOOF;
588 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
589 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl) | RCVCTL_ENABLE);
590 
591 	bcmeth_rxq_produce(sc, &sc->sc_rxq);	/* fill with rx buffers */
592 
593 #if 0
594 	aprint_normal_dev(sc->sc_dev,
595 	    "devctl=%#x ucmdcfg=%#x xmtctl=%#x rcvctl=%#x\n",
596 	    devctl, sc->sc_cmdcfg,
597 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl),
598 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl));
599 #endif
600 
601 	sc->sc_soft_flags = 0;
602 
603 	bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
604 
605 	ifp->if_flags |= IFF_RUNNING;
606 
607 	return error;
608 }
609 
610 static void
611 bcmeth_ifstop(struct ifnet *ifp, int disable)
612 {
613 	struct bcmeth_softc * const sc = ifp->if_softc;
614 	struct bcmeth_txqueue * const txq = &sc->sc_txq;
615 	struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
616 
617 	KASSERT(!cpu_intr_p());
618 
619 	sc->sc_soft_flags = 0;
620 	sc->sc_work_flags = 0;
621 
622 	/* Disable Rx processing */
623 	bcmeth_write_4(sc, rxq->rxq_reg_rcvctl,
624 	    bcmeth_read_4(sc, rxq->rxq_reg_rcvctl) & ~RCVCTL_ENABLE);
625 
626 	/* Disable Tx processing */
627 	bcmeth_write_4(sc, txq->txq_reg_xmtctl,
628 	    bcmeth_read_4(sc, txq->txq_reg_xmtctl) & ~XMTCTL_ENABLE);
629 
630 	/* Disable all interrupts */
631 	bcmeth_write_4(sc, GMAC_INTMASK, 0);
632 
633 	for (;;) {
634 		uint32_t tx0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
635 		uint32_t rx0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
636 		if (__SHIFTOUT(tx0, XMTSTATE) == XMTSTATE_DIS
637 		    && __SHIFTOUT(rx0, RCVSTATE) == RCVSTATE_DIS)
638 			break;
639 		delay(50);
640 	}
641 	/*
642 	 * Now reset the controller.
643 	 *
644 	 * 3. Set SW_RESET bit in UNIMAC_COMMAND_CONFIG register
645 	 * 4. Clear SW_RESET bit in UNIMAC_COMMAND_CONFIG register
646 	 */
647 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, SW_RESET);
648 	bcmeth_write_4(sc, GMAC_INTSTATUS, ~0);
649 	sc->sc_intmask = 0;
650 	ifp->if_flags &= ~IFF_RUNNING;
651 
652 	/*
653 	 * Let's consume any remaining transmitted packets.  And if we are
654 	 * disabling the interface, purge ourselves of any untransmitted
655 	 * packets.  But don't consume any received packets, just drop them.
656 	 * If we aren't disabling the interface, save the mbufs in the
657 	 * receive queue for reuse.
658 	 */
659 	bcmeth_rxq_purge(sc, &sc->sc_rxq, disable);
660 	bcmeth_txq_consume(sc, &sc->sc_txq);
661 	if (disable) {
662 		bcmeth_txq_purge(sc, &sc->sc_txq);
663 		IF_PURGE(&ifp->if_snd);
664 	}
665 
666 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, 0);
667 }
668 
669 static void
670 bcmeth_ifwatchdog(struct ifnet *ifp)
671 {
672 }
673 
674 static int
675 bcmeth_ifioctl(struct ifnet *ifp, u_long cmd, void *data)
676 {
677 	const int s = splnet();
678 	int error;
679 
680 	switch (cmd) {
681 	default:
682 		error = ether_ioctl(ifp, cmd, data);
683 		if (error != ENETRESET)
684 			break;
685 
686 		if (cmd == SIOCADDMULTI || cmd == SIOCDELMULTI) {
687 			error = 0;
688 			break;
689 		}
690 		error = bcmeth_ifinit(ifp);
691 		break;
692 	}
693 
694 	splx(s);
695 	return error;
696 }
697 
698 static void
699 bcmeth_rxq_desc_presync(
700 	struct bcmeth_softc *sc,
701 	struct bcmeth_rxqueue *rxq,
702 	struct gmac_rxdb *rxdb,
703 	size_t count)
704 {
705 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
706 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
707 	    BUS_DMASYNC_PREWRITE);
708 }
709 
710 static void
711 bcmeth_rxq_desc_postsync(
712 	struct bcmeth_softc *sc,
713 	struct bcmeth_rxqueue *rxq,
714 	struct gmac_rxdb *rxdb,
715 	size_t count)
716 {
717 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
718 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
719 	    BUS_DMASYNC_POSTWRITE);
720 }
721 
722 static void
723 bcmeth_txq_desc_presync(
724 	struct bcmeth_softc *sc,
725 	struct bcmeth_txqueue *txq,
726 	struct gmac_txdb *txdb,
727 	size_t count)
728 {
729 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
730 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
731 	    BUS_DMASYNC_PREWRITE);
732 }
733 
734 static void
735 bcmeth_txq_desc_postsync(
736 	struct bcmeth_softc *sc,
737 	struct bcmeth_txqueue *txq,
738 	struct gmac_txdb *txdb,
739 	size_t count)
740 {
741 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
742 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
743 	    BUS_DMASYNC_POSTWRITE);
744 }
745 
746 static bus_dmamap_t
747 bcmeth_mapcache_get(
748 	struct bcmeth_softc *sc,
749 	struct bcmeth_mapcache *dmc)
750 {
751 	KASSERT(dmc->dmc_nmaps > 0);
752 	KASSERT(dmc->dmc_maps[dmc->dmc_nmaps-1] != NULL);
753 	return dmc->dmc_maps[--dmc->dmc_nmaps];
754 }
755 
756 static void
757 bcmeth_mapcache_put(
758 	struct bcmeth_softc *sc,
759 	struct bcmeth_mapcache *dmc,
760 	bus_dmamap_t map)
761 {
762 	KASSERT(map != NULL);
763 	KASSERT(dmc->dmc_nmaps < dmc->dmc_maxmaps);
764 	dmc->dmc_maps[dmc->dmc_nmaps++] = map;
765 }
766 
767 static void
768 bcmeth_mapcache_destroy(
769 	struct bcmeth_softc *sc,
770 	struct bcmeth_mapcache *dmc)
771 {
772 	const size_t dmc_size =
773 	    offsetof(struct bcmeth_mapcache, dmc_maps[dmc->dmc_maxmaps]);
774 
775 	for (u_int i = 0; i < dmc->dmc_maxmaps; i++) {
776 		bus_dmamap_destroy(sc->sc_dmat, dmc->dmc_maps[i]);
777 	}
778 	kmem_intr_free(dmc, dmc_size);
779 }
780 
781 static int
782 bcmeth_mapcache_create(
783 	struct bcmeth_softc *sc,
784 	struct bcmeth_mapcache **dmc_p,
785 	size_t maxmaps,
786 	size_t maxmapsize,
787 	size_t maxseg)
788 {
789 	const size_t dmc_size =
790 	    offsetof(struct bcmeth_mapcache, dmc_maps[maxmaps]);
791 	struct bcmeth_mapcache * const dmc =
792 		kmem_intr_zalloc(dmc_size, KM_NOSLEEP);
793 
794 	dmc->dmc_maxmaps = maxmaps;
795 	dmc->dmc_nmaps = maxmaps;
796 	dmc->dmc_maxmapsize = maxmapsize;
797 	dmc->dmc_maxseg = maxseg;
798 
799 	for (u_int i = 0; i < maxmaps; i++) {
800 		int error = bus_dmamap_create(sc->sc_dmat, dmc->dmc_maxmapsize,
801 		     dmc->dmc_maxseg, dmc->dmc_maxmapsize, 0,
802 		     BUS_DMA_WAITOK | BUS_DMA_ALLOCNOW, &dmc->dmc_maps[i]);
803 		if (error) {
804 			aprint_error_dev(sc->sc_dev,
805 			    "failed to creat dma map cache "
806 			    "entry %u of %zu: %d\n",
807 			    i, maxmaps, error);
808 			while (i-- > 0) {
809 				bus_dmamap_destroy(sc->sc_dmat,
810 				    dmc->dmc_maps[i]);
811 			}
812 			kmem_intr_free(dmc, dmc_size);
813 			return error;
814 		}
815 		KASSERT(dmc->dmc_maps[i] != NULL);
816 	}
817 
818 	*dmc_p = dmc;
819 
820 	return 0;
821 }
822 
823 #if 0
824 static void
825 bcmeth_dmamem_free(
826 	bus_dma_tag_t dmat,
827 	size_t map_size,
828 	bus_dma_segment_t *seg,
829 	bus_dmamap_t map,
830 	void *kvap)
831 {
832 	bus_dmamap_destroy(dmat, map);
833 	bus_dmamem_unmap(dmat, kvap, map_size);
834 	bus_dmamem_free(dmat, seg, 1);
835 }
836 #endif
837 
838 static int
839 bcmeth_dmamem_alloc(
840 	bus_dma_tag_t dmat,
841 	size_t map_size,
842 	bus_dma_segment_t *seg,
843 	bus_dmamap_t *map,
844 	void **kvap)
845 {
846 	int error;
847 	int nseg;
848 
849 	*kvap = NULL;
850 	*map = NULL;
851 
852 	error = bus_dmamem_alloc(dmat, map_size, 2*PAGE_SIZE, 0,
853 	   seg, 1, &nseg, 0);
854 	if (error)
855 		return error;
856 
857 	KASSERT(nseg == 1);
858 
859 	error = bus_dmamem_map(dmat, seg, nseg, map_size, (void **)kvap, 0);
860 	if (error == 0) {
861 		error = bus_dmamap_create(dmat, map_size, 1, map_size, 0, 0,
862 		    map);
863 		if (error == 0) {
864 			error = bus_dmamap_load(dmat, *map, *kvap, map_size,
865 			    NULL, 0);
866 			if (error == 0)
867 				return 0;
868 			bus_dmamap_destroy(dmat, *map);
869 			*map = NULL;
870 		}
871 		bus_dmamem_unmap(dmat, *kvap, map_size);
872 		*kvap = NULL;
873 	}
874 	bus_dmamem_free(dmat, seg, nseg);
875 	return 0;
876 }
877 
878 static struct mbuf *
879 bcmeth_rx_buf_alloc(
880 	struct bcmeth_softc *sc)
881 {
882 	struct mbuf *m = m_gethdr(M_DONTWAIT, MT_DATA);
883 	if (m == NULL) {
884 		printf("%s:%d: %s\n", __func__, __LINE__, "m_gethdr");
885 		return NULL;
886 	}
887 	MCLGET(m, M_DONTWAIT);
888 	if ((m->m_flags & M_EXT) == 0) {
889 		printf("%s:%d: %s\n", __func__, __LINE__, "MCLGET");
890 		m_freem(m);
891 		return NULL;
892 	}
893 	m->m_len = m->m_pkthdr.len = m->m_ext.ext_size;
894 
895 	bus_dmamap_t map = bcmeth_mapcache_get(sc, sc->sc_rx_mapcache);
896 	if (map == NULL) {
897 		printf("%s:%d: %s\n", __func__, __LINE__, "map get");
898 		m_freem(m);
899 		return NULL;
900 	}
901 	M_SETCTX(m, map);
902 	m->m_len = m->m_pkthdr.len = MCLBYTES;
903 	int error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
904 	    BUS_DMA_READ | BUS_DMA_NOWAIT);
905 	if (error) {
906 		aprint_error_dev(sc->sc_dev, "fail to load rx dmamap: %d\n",
907 		    error);
908 		M_SETCTX(m, NULL);
909 		m_freem(m);
910 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
911 		return NULL;
912 	}
913 	KASSERT(map->dm_mapsize == MCLBYTES);
914 #ifdef BCMETH_RCVMAGIC
915 	*mtod(m, uint32_t *) = htole32(BCMETH_RCVMAGIC);
916 	bus_dmamap_sync(sc->sc_dmat, map, 0, sizeof(uint32_t),
917 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
918 	bus_dmamap_sync(sc->sc_dmat, map, sizeof(uint32_t),
919 	    map->dm_mapsize - sizeof(uint32_t), BUS_DMASYNC_PREREAD);
920 #else
921 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
922 	    BUS_DMASYNC_PREREAD);
923 #endif
924 
925 	return m;
926 }
927 
928 static void
929 bcmeth_rx_map_unload(
930 	struct bcmeth_softc *sc,
931 	struct mbuf *m)
932 {
933 	KASSERT(m);
934 	for (; m != NULL; m = m->m_next) {
935 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
936 		KASSERT(map);
937 		KASSERT(map->dm_mapsize == MCLBYTES);
938 		bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_len,
939 		    BUS_DMASYNC_POSTREAD);
940 		bus_dmamap_unload(sc->sc_dmat, map);
941 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
942 		M_SETCTX(m, NULL);
943 	}
944 }
945 
946 static bool
947 bcmeth_rxq_produce(
948 	struct bcmeth_softc *sc,
949 	struct bcmeth_rxqueue *rxq)
950 {
951 	struct gmac_rxdb *producer = rxq->rxq_producer;
952 	bool produced = false;
953 
954 	while (rxq->rxq_inuse < rxq->rxq_threshold) {
955 		struct mbuf *m;
956 		IF_DEQUEUE(&sc->sc_rx_bufcache, m);
957 		if (m == NULL) {
958 			m = bcmeth_rx_buf_alloc(sc);
959 			if (m == NULL) {
960 				printf("%s: bcmeth_rx_buf_alloc failed\n",
961 				    __func__);
962 				break;
963 			}
964 		}
965 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
966 		KASSERT(map);
967 
968 		producer->rxdb_buflen = htole32(MCLBYTES);
969 		producer->rxdb_addrlo = htole32(map->dm_segs[0].ds_addr);
970 		producer->rxdb_flags &= htole32(RXDB_FLAG_ET);
971 		*rxq->rxq_mtail = m;
972 		rxq->rxq_mtail = &m->m_next;
973 		m->m_len = MCLBYTES;
974 		m->m_next = NULL;
975 		rxq->rxq_inuse++;
976 		if (++producer == rxq->rxq_last) {
977 			membar_producer();
978 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
979 			    rxq->rxq_last - rxq->rxq_producer);
980 			producer = rxq->rxq_producer = rxq->rxq_first;
981 		}
982 		produced = true;
983 	}
984 	if (produced) {
985 		membar_producer();
986 		if (producer != rxq->rxq_producer) {
987 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
988 			    producer - rxq->rxq_producer);
989 			rxq->rxq_producer = producer;
990 		}
991 		bcmeth_write_4(sc, rxq->rxq_reg_rcvptr,
992 		    rxq->rxq_descmap->dm_segs[0].ds_addr
993 		    + ((uintptr_t)producer & RCVPTR));
994 	}
995 	return true;
996 }
997 
998 static void
999 bcmeth_rx_input(
1000 	struct bcmeth_softc *sc,
1001 	struct mbuf *m,
1002 	uint32_t rxdb_flags)
1003 {
1004 	struct ifnet * const ifp = &sc->sc_if;
1005 
1006 	bcmeth_rx_map_unload(sc, m);
1007 
1008 	m_adj(m, sc->sc_rcvoffset);
1009 
1010 	/*
1011 	 * If we are in promiscuous mode and this isn't a multicast, check the
1012 	 * destination address to make sure it matches our own.  If it doesn't,
1013 	 * mark the packet as being received promiscuously.
1014 	 */
1015 	if ((sc->sc_cmdcfg & PROMISC_EN)
1016 	    && (m->m_data[0] & 1) == 0
1017 	    && (*(uint16_t *)&m->m_data[0] != sc->sc_macaddr[0]
1018 		|| *(uint32_t *)&m->m_data[2] != sc->sc_macaddr[1])) {
1019 		m->m_flags |= M_PROMISC;
1020 	}
1021 	m_set_rcvif(m, ifp);
1022 
1023 	/*
1024 	 * Let's give it to the network subsystm to deal with.
1025 	 */
1026 #ifdef BCMETH_MPSAFE
1027 	mutex_exit(sc->sc_lock);
1028 	if_input(ifp, m);
1029 	mutex_enter(sc->sc_lock);
1030 #else
1031 	int s = splnet();
1032 	if_input(ifp, m);
1033 	splx(s);
1034 #endif
1035 }
1036 
1037 static bool
1038 bcmeth_rxq_consume(
1039 	struct bcmeth_softc *sc,
1040 	struct bcmeth_rxqueue *rxq,
1041 	size_t atmost)
1042 {
1043 	struct ifnet * const ifp = &sc->sc_if;
1044 	struct gmac_rxdb *consumer = rxq->rxq_consumer;
1045 	size_t rxconsumed = 0;
1046 	bool didconsume = false;
1047 
1048 	while (atmost-- > 0) {
1049 		if (consumer == rxq->rxq_producer) {
1050 			KASSERT(rxq->rxq_inuse == 0);
1051 			break;
1052 		}
1053 
1054 		uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1055 		uint32_t currdscr = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1056 		if (consumer == rxq->rxq_first + currdscr) {
1057 			break;
1058 		}
1059 		bcmeth_rxq_desc_postsync(sc, rxq, consumer, 1);
1060 
1061 		/*
1062 		 * We own this packet again.  Copy the rxsts word from it.
1063 		 */
1064 		rxconsumed++;
1065 		didconsume = true;
1066 		uint32_t rxsts;
1067 		KASSERT(rxq->rxq_mhead != NULL);
1068 		bus_dmamap_t map = M_GETCTX(rxq->rxq_mhead, bus_dmamap_t);
1069 		bus_dmamap_sync(sc->sc_dmat, map, 0, arm_dcache_align,
1070 		    BUS_DMASYNC_POSTREAD);
1071 		memcpy(&rxsts, rxq->rxq_mhead->m_data, 4);
1072 		rxsts = le32toh(rxsts);
1073 #if 0
1074 		KASSERTMSG(rxsts != BCMETH_RCVMAGIC, "currdscr=%u consumer=%zd",
1075 		    currdscr, consumer - rxq->rxq_first);
1076 #endif
1077 
1078 		/*
1079 		 * Get the count of descriptors.  Fetch the correct number
1080 		 * of mbufs.
1081 		 */
1082 #ifdef BCMETH_RCVMAGIC
1083 		size_t desc_count = rxsts != BCMETH_RCVMAGIC
1084 		    ? __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1 : 1;
1085 #else
1086 		size_t desc_count = __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1;
1087 #endif
1088 		struct mbuf *m = rxq->rxq_mhead;
1089 		struct mbuf *m_last = m;
1090 		for (size_t i = 1; i < desc_count; i++) {
1091 			if (++consumer == rxq->rxq_last) {
1092 				consumer = rxq->rxq_first;
1093 			}
1094 			KASSERTMSG(consumer != rxq->rxq_first + currdscr,
1095 			    "i=%zu rxsts=%#x desc_count=%zu currdscr=%u "
1096 			    "consumer=%zd", i, rxsts, desc_count, currdscr,
1097 			    consumer - rxq->rxq_first);
1098 			m_last = m_last->m_next;
1099 		}
1100 
1101 		/*
1102 		 * Now remove it/them from the list of enqueued mbufs.
1103 		 */
1104 		if ((rxq->rxq_mhead = m_last->m_next) == NULL)
1105 			rxq->rxq_mtail = &rxq->rxq_mhead;
1106 		m_last->m_next = NULL;
1107 
1108 #ifdef BCMETH_RCVMAGIC
1109 		if (rxsts == BCMETH_RCVMAGIC) {
1110 			if_statinc(ifp, if_ierrors);
1111 			if ((m->m_ext.ext_paddr >> 28) == 8) {
1112 				BCMETH_EVCNT_INCR(sc->sc_ev_rx_badmagic_lo);
1113 			} else {
1114 				BCMETH_EVCNT_INCR( sc->sc_ev_rx_badmagic_hi);
1115 			}
1116 			IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1117 		} else
1118 #endif /* BCMETH_RCVMAGIC */
1119 		if (rxsts
1120 		    & (RXSTS_CRC_ERROR |RXSTS_OVERSIZED |RXSTS_PKT_OVERFLOW)) {
1121 			aprint_error_dev(sc->sc_dev,
1122 			    "[%zu]: count=%zu rxsts=%#x\n",
1123 			    consumer - rxq->rxq_first, desc_count, rxsts);
1124 			/*
1125 			 * We encountered an error, take the mbufs and add them
1126 			 * to the rx bufcache so we can quickly reuse them.
1127 			 */
1128 			if_statinc(ifp, if_ierrors);
1129 			do {
1130 				struct mbuf *m0 = m->m_next;
1131 				m->m_next = NULL;
1132 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1133 				m = m0;
1134 			} while (m);
1135 		} else {
1136 			uint32_t framelen = __SHIFTOUT(rxsts, RXSTS_FRAMELEN);
1137 			framelen += sc->sc_rcvoffset;
1138 			m->m_pkthdr.len = framelen;
1139 			if (desc_count == 1) {
1140 				KASSERT(framelen <= MCLBYTES);
1141 				m->m_len = framelen;
1142 			} else {
1143 				m_last->m_len = framelen & (MCLBYTES - 1);
1144 			}
1145 
1146 #ifdef BCMETH_MPSAFE
1147 			/*
1148 			 * Wrap at the last entry!
1149 			 */
1150 			if (++consumer == rxq->rxq_last) {
1151 				KASSERT(consumer[-1].rxdb_flags
1152 				    & htole32(RXDB_FLAG_ET));
1153 				rxq->rxq_consumer = rxq->rxq_first;
1154 			} else {
1155 				rxq->rxq_consumer = consumer;
1156 			}
1157 			rxq->rxq_inuse -= rxconsumed;
1158 #endif /* BCMETH_MPSAFE */
1159 
1160 			/*
1161 			 * Receive the packet (which releases our lock)
1162 			 */
1163 			bcmeth_rx_input(sc, m, rxsts);
1164 
1165 #ifdef BCMETH_MPSAFE
1166 			/*
1167 			 * Since we had to give up our lock, we need to
1168 			 * refresh these.
1169 			 */
1170 			consumer = rxq->rxq_consumer;
1171 			rxconsumed = 0;
1172 			continue;
1173 #endif /* BCMETH_MPSAFE */
1174 		}
1175 
1176 		/*
1177 		 * Wrap at the last entry!
1178 		 */
1179 		if (++consumer == rxq->rxq_last) {
1180 			KASSERT(consumer[-1].rxdb_flags & htole32(RXDB_FLAG_ET));
1181 			consumer = rxq->rxq_first;
1182 		}
1183 	}
1184 
1185 	/*
1186 	 * Update queue info.
1187 	 */
1188 	rxq->rxq_consumer = consumer;
1189 	rxq->rxq_inuse -= rxconsumed;
1190 
1191 	/*
1192 	 * Did we consume anything?
1193 	 */
1194 	return didconsume;
1195 }
1196 
1197 static void
1198 bcmeth_rxq_purge(
1199 	struct bcmeth_softc *sc,
1200 	struct bcmeth_rxqueue *rxq,
1201 	bool discard)
1202 {
1203 	struct mbuf *m;
1204 
1205 	if ((m = rxq->rxq_mhead) != NULL) {
1206 		if (discard) {
1207 			bcmeth_rx_map_unload(sc, m);
1208 			m_freem(m);
1209 		} else {
1210 			while (m != NULL) {
1211 				struct mbuf *m0 = m->m_next;
1212 				m->m_next = NULL;
1213 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1214 				m = m0;
1215 			}
1216 		}
1217 	}
1218 
1219 	rxq->rxq_mhead = NULL;
1220 	rxq->rxq_mtail = &rxq->rxq_mhead;
1221 	rxq->rxq_inuse = 0;
1222 }
1223 
1224 static void
1225 bcmeth_rxq_reset(
1226 	struct bcmeth_softc *sc,
1227 	struct bcmeth_rxqueue *rxq)
1228 {
1229 	/*
1230 	 * sync all the descriptors
1231 	 */
1232 	bcmeth_rxq_desc_postsync(sc, rxq, rxq->rxq_first,
1233 	    rxq->rxq_last - rxq->rxq_first);
1234 
1235 	/*
1236 	 * Make sure we own all descriptors in the ring.
1237 	 */
1238 	struct gmac_rxdb *rxdb;
1239 	for (rxdb = rxq->rxq_first; rxdb < rxq->rxq_last - 1; rxdb++) {
1240 		rxdb->rxdb_flags = htole32(RXDB_FLAG_IC);
1241 	}
1242 
1243 	/*
1244 	 * Last descriptor has the wrap flag.
1245 	 */
1246 	rxdb->rxdb_flags = htole32(RXDB_FLAG_ET | RXDB_FLAG_IC);
1247 
1248 	/*
1249 	 * Reset the producer consumer indexes.
1250 	 */
1251 	rxq->rxq_consumer = rxq->rxq_first;
1252 	rxq->rxq_producer = rxq->rxq_first;
1253 	rxq->rxq_inuse = 0;
1254 	if (rxq->rxq_threshold < BCMETH_MINRXMBUFS)
1255 		rxq->rxq_threshold = BCMETH_MINRXMBUFS;
1256 
1257 	sc->sc_intmask |= RCVINT | RCVFIFOOF | RCVDESCUF;
1258 
1259 	/*
1260 	 * Restart the receiver at the first descriptor
1261 	 */
1262 	bcmeth_write_4(sc, rxq->rxq_reg_rcvaddrlo,
1263 	    rxq->rxq_descmap->dm_segs[0].ds_addr);
1264 }
1265 
1266 static int
1267 bcmeth_rxq_attach(
1268 	struct bcmeth_softc *sc,
1269 	struct bcmeth_rxqueue *rxq,
1270 	u_int qno)
1271 {
1272 	size_t desc_count = BCMETH_RINGSIZE / sizeof(rxq->rxq_first[0]);
1273 	int error;
1274 	void *descs;
1275 
1276 	KASSERT(desc_count == 256 || desc_count == 512);
1277 
1278 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1279 	   &rxq->rxq_descmap_seg, &rxq->rxq_descmap, &descs);
1280 	if (error)
1281 		return error;
1282 
1283 	memset(descs, 0, BCMETH_RINGSIZE);
1284 	rxq->rxq_first = descs;
1285 	rxq->rxq_last = rxq->rxq_first + desc_count;
1286 	rxq->rxq_consumer = descs;
1287 	rxq->rxq_producer = descs;
1288 
1289 	bcmeth_rxq_purge(sc, rxq, true);
1290 	bcmeth_rxq_reset(sc, rxq);
1291 
1292 	rxq->rxq_reg_rcvaddrlo = GMAC_RCVADDR_LOW;
1293 	rxq->rxq_reg_rcvctl = GMAC_RCVCONTROL;
1294 	rxq->rxq_reg_rcvptr = GMAC_RCVPTR;
1295 	rxq->rxq_reg_rcvsts0 = GMAC_RCVSTATUS0;
1296 	rxq->rxq_reg_rcvsts1 = GMAC_RCVSTATUS1;
1297 
1298 	return 0;
1299 }
1300 
1301 static bool
1302 bcmeth_txq_active_p(
1303 	struct bcmeth_softc * const sc,
1304 	struct bcmeth_txqueue *txq)
1305 {
1306 	return !IF_IS_EMPTY(&txq->txq_mbufs);
1307 }
1308 
1309 static bool
1310 bcmeth_txq_fillable_p(
1311 	struct bcmeth_softc * const sc,
1312 	struct bcmeth_txqueue *txq)
1313 {
1314 	return txq->txq_free >= txq->txq_threshold;
1315 }
1316 
1317 static int
1318 bcmeth_txq_attach(
1319 	struct bcmeth_softc *sc,
1320 	struct bcmeth_txqueue *txq,
1321 	u_int qno)
1322 {
1323 	size_t desc_count = BCMETH_RINGSIZE / sizeof(txq->txq_first[0]);
1324 	int error;
1325 	void *descs;
1326 
1327 	KASSERT(desc_count == 256 || desc_count == 512);
1328 
1329 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1330 	   &txq->txq_descmap_seg, &txq->txq_descmap, &descs);
1331 	if (error)
1332 		return error;
1333 
1334 	memset(descs, 0, BCMETH_RINGSIZE);
1335 	txq->txq_first = descs;
1336 	txq->txq_last = txq->txq_first + desc_count;
1337 	txq->txq_consumer = descs;
1338 	txq->txq_producer = descs;
1339 
1340 	IFQ_SET_MAXLEN(&txq->txq_mbufs, BCMETH_MAXTXMBUFS);
1341 
1342 	txq->txq_reg_xmtaddrlo = GMAC_XMTADDR_LOW;
1343 	txq->txq_reg_xmtctl = GMAC_XMTCONTROL;
1344 	txq->txq_reg_xmtptr = GMAC_XMTPTR;
1345 	txq->txq_reg_xmtsts0 = GMAC_XMTSTATUS0;
1346 	txq->txq_reg_xmtsts1 = GMAC_XMTSTATUS1;
1347 
1348 	bcmeth_txq_reset(sc, txq);
1349 
1350 	return 0;
1351 }
1352 
1353 static int
1354 bcmeth_txq_map_load(
1355 	struct bcmeth_softc *sc,
1356 	struct bcmeth_txqueue *txq,
1357 	struct mbuf *m)
1358 {
1359 	bus_dmamap_t map;
1360 	int error;
1361 
1362 	map = M_GETCTX(m, bus_dmamap_t);
1363 	if (map != NULL)
1364 		return 0;
1365 
1366 	map = bcmeth_mapcache_get(sc, sc->sc_tx_mapcache);
1367 	if (map == NULL)
1368 		return ENOMEM;
1369 
1370 	error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
1371 	    BUS_DMA_WRITE | BUS_DMA_NOWAIT);
1372 	if (error)
1373 		return error;
1374 
1375 	bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_pkthdr.len,
1376 	    BUS_DMASYNC_PREWRITE);
1377 	M_SETCTX(m, map);
1378 	return 0;
1379 }
1380 
1381 static void
1382 bcmeth_txq_map_unload(
1383 	struct bcmeth_softc *sc,
1384 	struct bcmeth_txqueue *txq,
1385 	struct mbuf *m)
1386 {
1387 	KASSERT(m);
1388 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1389 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1390 	    BUS_DMASYNC_POSTWRITE);
1391 	bus_dmamap_unload(sc->sc_dmat, map);
1392 	bcmeth_mapcache_put(sc, sc->sc_tx_mapcache, map);
1393 }
1394 
1395 static bool
1396 bcmeth_txq_produce(
1397 	struct bcmeth_softc *sc,
1398 	struct bcmeth_txqueue *txq,
1399 	struct mbuf *m)
1400 {
1401 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1402 
1403 	if (map->dm_nsegs > txq->txq_free)
1404 		return false;
1405 
1406 	/*
1407 	 * TCP Offload flag must be set in the first descriptor.
1408 	 */
1409 	struct gmac_txdb *producer = txq->txq_producer;
1410 	uint32_t first_flags = TXDB_FLAG_SF;
1411 	uint32_t last_flags = TXDB_FLAG_EF;
1412 
1413 	/*
1414 	 * If we've produced enough descriptors without consuming any
1415 	 * we need to ask for an interrupt to reclaim some.
1416 	 */
1417 	txq->txq_lastintr += map->dm_nsegs;
1418 	if (txq->txq_lastintr >= txq->txq_threshold
1419 	    || txq->txq_mbufs.ifq_len + 1 == txq->txq_mbufs.ifq_maxlen) {
1420 		txq->txq_lastintr = 0;
1421 		last_flags |= TXDB_FLAG_IC;
1422 	}
1423 
1424 	KASSERT(producer != txq->txq_last);
1425 
1426 	struct gmac_txdb *start = producer;
1427 	size_t count = map->dm_nsegs;
1428 	producer->txdb_flags |= htole32(first_flags);
1429 	producer->txdb_addrlo = htole32(map->dm_segs[0].ds_addr);
1430 	producer->txdb_buflen = htole32(map->dm_segs[0].ds_len);
1431 	for (u_int i = 1; i < map->dm_nsegs; i++) {
1432 #if 0
1433 		printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1434 		    le32toh(producer->txdb_flags),
1435 		    le32toh(producer->txdb_buflen),
1436 		    le32toh(producer->txdb_addrlo),
1437 		    le32toh(producer->txdb_addrhi));
1438 #endif
1439 		if (__predict_false(++producer == txq->txq_last)) {
1440 			bcmeth_txq_desc_presync(sc, txq, start,
1441 			    txq->txq_last - start);
1442 			count -= txq->txq_last - start;
1443 			producer = txq->txq_first;
1444 			start = txq->txq_first;
1445 		}
1446 		producer->txdb_addrlo = htole32(map->dm_segs[i].ds_addr);
1447 		producer->txdb_buflen = htole32(map->dm_segs[i].ds_len);
1448 	}
1449 	producer->txdb_flags |= htole32(last_flags);
1450 #if 0
1451 	printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1452 	    le32toh(producer->txdb_flags), le32toh(producer->txdb_buflen),
1453 	    le32toh(producer->txdb_addrlo), le32toh(producer->txdb_addrhi));
1454 #endif
1455 	if (count)
1456 		bcmeth_txq_desc_presync(sc, txq, start, count);
1457 
1458 	/*
1459 	 * Reduce free count by the number of segments we consumed.
1460 	 */
1461 	txq->txq_free -= map->dm_nsegs;
1462 	KASSERT(map->dm_nsegs == 1 || txq->txq_producer != producer);
1463 	KASSERT(map->dm_nsegs == 1
1464 	    || (txq->txq_producer->txdb_flags & htole32(TXDB_FLAG_EF)) == 0);
1465 	KASSERT(producer->txdb_flags & htole32(TXDB_FLAG_EF));
1466 
1467 #if 0
1468 	printf("%s: mbuf %p: produced a %u byte packet in %u segments "
1469 	    "(%zd..%zd)\n", __func__, m, m->m_pkthdr.len, map->dm_nsegs,
1470 	    txq->txq_producer - txq->txq_first, producer - txq->txq_first);
1471 #endif
1472 
1473 	if (producer + 1 == txq->txq_last)
1474 		txq->txq_producer = txq->txq_first;
1475 	else
1476 		txq->txq_producer = producer + 1;
1477 	IF_ENQUEUE(&txq->txq_mbufs, m);
1478 
1479 	/*
1480 	 * Let the transmitter know there's more to do
1481 	 */
1482 	bcmeth_write_4(sc, txq->txq_reg_xmtptr,
1483 	    txq->txq_descmap->dm_segs[0].ds_addr
1484 	    + ((uintptr_t)txq->txq_producer & XMT_LASTDSCR));
1485 
1486 	return true;
1487 }
1488 
1489 static struct mbuf *
1490 bcmeth_copy_packet(struct mbuf *m)
1491 {
1492 	struct mbuf *mext = NULL;
1493 	size_t misalignment = 0;
1494 	size_t hlen = 0;
1495 
1496 	for (mext = m; mext != NULL; mext = mext->m_next) {
1497 		if (mext->m_flags & M_EXT) {
1498 			misalignment = mtod(mext, vaddr_t) & arm_dcache_align;
1499 			break;
1500 		}
1501 		hlen += m->m_len;
1502 	}
1503 
1504 	struct mbuf *n = m->m_next;
1505 	if (m != mext && hlen + misalignment <= MHLEN && false) {
1506 		KASSERT(m->m_pktdat <= m->m_data
1507 		    && m->m_data <= &m->m_pktdat[MHLEN - m->m_len]);
1508 		size_t oldoff = m->m_data - m->m_pktdat;
1509 		size_t off;
1510 		if (mext == NULL) {
1511 			off = (oldoff + hlen > MHLEN) ? 0 : oldoff;
1512 		} else {
1513 			off = MHLEN - (hlen + misalignment);
1514 		}
1515 		KASSERT(off + hlen + misalignment <= MHLEN);
1516 		if (((oldoff ^ off) & arm_dcache_align) != 0 || off < oldoff) {
1517 			memmove(&m->m_pktdat[off], m->m_data, m->m_len);
1518 			m->m_data = &m->m_pktdat[off];
1519 		}
1520 		m_copydata(n, 0, hlen - m->m_len, &m->m_data[m->m_len]);
1521 		m->m_len = hlen;
1522 		m->m_next = mext;
1523 		while (n != mext) {
1524 			n = m_free(n);
1525 		}
1526 		return m;
1527 	}
1528 
1529 	struct mbuf *m0 = m_gethdr(M_DONTWAIT, m->m_type);
1530 	if (m0 == NULL) {
1531 		return NULL;
1532 	}
1533 	m_copy_pkthdr(m0, m);
1534 	MCLAIM(m0, m->m_owner);
1535 	if (m0->m_pkthdr.len > MHLEN) {
1536 		MCLGET(m0, M_DONTWAIT);
1537 		if ((m0->m_flags & M_EXT) == 0) {
1538 			m_freem(m0);
1539 			return NULL;
1540 		}
1541 	}
1542 	m0->m_len = m->m_pkthdr.len;
1543 	m_copydata(m, 0, m0->m_len, mtod(m0, void *));
1544 	m_freem(m);
1545 	return m0;
1546 }
1547 
1548 static bool
1549 bcmeth_txq_enqueue(
1550 	struct bcmeth_softc *sc,
1551 	struct bcmeth_txqueue *txq)
1552 {
1553 	for (;;) {
1554 		if (IF_QFULL(&txq->txq_mbufs))
1555 			return false;
1556 		struct mbuf *m = txq->txq_next;
1557 		if (m == NULL) {
1558 			int s = splnet();
1559 			IF_DEQUEUE(&sc->sc_if.if_snd, m);
1560 			splx(s);
1561 			if (m == NULL)
1562 				return true;
1563 			M_SETCTX(m, NULL);
1564 		} else {
1565 			txq->txq_next = NULL;
1566 		}
1567 		/*
1568 		 * If LINK2 is set and this packet uses multiple mbufs,
1569 		 * consolidate it into a single mbuf.
1570 		 */
1571 		if (m->m_next != NULL && (sc->sc_if.if_flags & IFF_LINK2)) {
1572 			struct mbuf *m0 = bcmeth_copy_packet(m);
1573 			if (m0 == NULL) {
1574 				txq->txq_next = m;
1575 				return true;
1576 			}
1577 			m = m0;
1578 		}
1579 		int error = bcmeth_txq_map_load(sc, txq, m);
1580 		if (error) {
1581 			aprint_error_dev(sc->sc_dev,
1582 			    "discarded packet due to "
1583 			    "dmamap load failure: %d\n", error);
1584 			m_freem(m);
1585 			continue;
1586 		}
1587 		KASSERT(txq->txq_next == NULL);
1588 		if (!bcmeth_txq_produce(sc, txq, m)) {
1589 			txq->txq_next = m;
1590 			return false;
1591 		}
1592 		KASSERT(txq->txq_next == NULL);
1593 	}
1594 }
1595 
1596 static bool
1597 bcmeth_txq_consume(
1598 	struct bcmeth_softc *sc,
1599 	struct bcmeth_txqueue *txq)
1600 {
1601 	struct ifnet * const ifp = &sc->sc_if;
1602 	struct gmac_txdb *consumer = txq->txq_consumer;
1603 	size_t txfree = 0;
1604 
1605 #if 0
1606 	printf("%s: entry: free=%zu\n", __func__, txq->txq_free);
1607 #endif
1608 
1609 	for (;;) {
1610 		if (consumer == txq->txq_producer) {
1611 			txq->txq_consumer = consumer;
1612 			txq->txq_free += txfree;
1613 			txq->txq_lastintr -= uimin(txq->txq_lastintr, txfree);
1614 #if 0
1615 			printf("%s: empty: freed %zu descriptors going from "
1616 			    "%zu to %zu\n", __func__, txfree,
1617 			    txq->txq_free - txfree, txq->txq_free);
1618 #endif
1619 			KASSERT(txq->txq_lastintr == 0);
1620 			KASSERT(txq->txq_free
1621 			    == txq->txq_last - txq->txq_first - 1);
1622 			return true;
1623 		}
1624 		bcmeth_txq_desc_postsync(sc, txq, consumer, 1);
1625 		uint32_t s0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
1626 		if (consumer == txq->txq_first + __SHIFTOUT(s0, XMT_CURRDSCR)) {
1627 			txq->txq_consumer = consumer;
1628 			txq->txq_free += txfree;
1629 			txq->txq_lastintr -= uimin(txq->txq_lastintr, txfree);
1630 #if 0
1631 			printf("%s: freed %zu descriptors\n",
1632 			    __func__, txfree);
1633 #endif
1634 			return bcmeth_txq_fillable_p(sc, txq);
1635 		}
1636 
1637 		/*
1638 		 * If this is the last descriptor in the chain, get the
1639 		 * mbuf, free its dmamap, and free the mbuf chain itself.
1640 		 */
1641 		const uint32_t txdb_flags = le32toh(consumer->txdb_flags);
1642 		if (txdb_flags & TXDB_FLAG_EF) {
1643 			struct mbuf *m;
1644 
1645 			IF_DEQUEUE(&txq->txq_mbufs, m);
1646 			KASSERT(m);
1647 			bcmeth_txq_map_unload(sc, txq, m);
1648 #if 0
1649 			printf("%s: mbuf %p: consumed a %u byte packet\n",
1650 			    __func__, m, m->m_pkthdr.len);
1651 #endif
1652 			bpf_mtap(ifp, m, BPF_D_OUT);
1653 			if_statinc(ifp, if_opackets);
1654 			if_statadd(ifp, if_obytes,  m->m_pkthdr.len);
1655 			if (m->m_flags & M_MCAST)
1656 				if_statinc(ifp, if_omcasts);
1657 			m_freem(m);
1658 		}
1659 
1660 		/*
1661 		 * We own this packet again.  Clear all flags except wrap.
1662 		 */
1663 		txfree++;
1664 
1665 		/*
1666 		 * Wrap at the last entry!
1667 		 */
1668 		if (txdb_flags & TXDB_FLAG_ET) {
1669 			consumer->txdb_flags = htole32(TXDB_FLAG_ET);
1670 			KASSERT(consumer + 1 == txq->txq_last);
1671 			consumer = txq->txq_first;
1672 		} else {
1673 			consumer->txdb_flags = 0;
1674 			consumer++;
1675 			KASSERT(consumer < txq->txq_last);
1676 		}
1677 	}
1678 }
1679 
1680 static void
1681 bcmeth_txq_purge(
1682 	struct bcmeth_softc *sc,
1683 	struct bcmeth_txqueue *txq)
1684 {
1685 	struct mbuf *m;
1686 	KASSERT((bcmeth_read_4(sc, UNIMAC_COMMAND_CONFIG) & TX_ENA) == 0);
1687 
1688 	for (;;) {
1689 		IF_DEQUEUE(&txq->txq_mbufs, m);
1690 		if (m == NULL)
1691 			break;
1692 		bcmeth_txq_map_unload(sc, txq, m);
1693 		m_freem(m);
1694 	}
1695 	if ((m = txq->txq_next) != NULL) {
1696 		txq->txq_next = NULL;
1697 		bcmeth_txq_map_unload(sc, txq, m);
1698 		m_freem(m);
1699 	}
1700 }
1701 
1702 static void
1703 bcmeth_txq_reset(
1704 	struct bcmeth_softc *sc,
1705 	struct bcmeth_txqueue *txq)
1706 {
1707 	/*
1708 	 * sync all the descriptors
1709 	 */
1710 	bcmeth_txq_desc_postsync(sc, txq, txq->txq_first,
1711 	    txq->txq_last - txq->txq_first);
1712 
1713 	/*
1714 	 * Make sure we own all descriptors in the ring.
1715 	 */
1716 	struct gmac_txdb *txdb;
1717 	for (txdb = txq->txq_first; txdb < txq->txq_last - 1; txdb++) {
1718 		txdb->txdb_flags = 0;
1719 	}
1720 
1721 	/*
1722 	 * Last descriptor has the wrap flag.
1723 	 */
1724 	txdb->txdb_flags = htole32(TXDB_FLAG_ET);
1725 
1726 	/*
1727 	 * Reset the producer consumer indexes.
1728 	 */
1729 	txq->txq_consumer = txq->txq_first;
1730 	txq->txq_producer = txq->txq_first;
1731 	txq->txq_free = txq->txq_last - txq->txq_first - 1;
1732 	txq->txq_threshold = txq->txq_free / 2;
1733 	txq->txq_lastintr = 0;
1734 
1735 	/*
1736 	 * What do we want to get interrupted on?
1737 	 */
1738 	sc->sc_intmask |= XMTINT_0 | XMTUF;
1739 
1740 	/*
1741 	 * Restart the transmitter at the first descriptor
1742 	 */
1743 	bcmeth_write_4(sc, txq->txq_reg_xmtaddrlo,
1744 	    txq->txq_descmap->dm_segs->ds_addr);
1745 }
1746 
1747 static void
1748 bcmeth_ifstart(struct ifnet *ifp)
1749 {
1750 	struct bcmeth_softc * const sc = ifp->if_softc;
1751 
1752 	if (__predict_false((ifp->if_flags & IFF_RUNNING) == 0)) {
1753 		return;
1754 	}
1755 
1756 #ifdef BCMETH_MPSAFETX
1757 	if (cpu_intr_p()) {
1758 #endif
1759 		atomic_or_uint(&sc->sc_soft_flags, SOFT_TXINTR);
1760 		softint_schedule(sc->sc_soft_ih);
1761 #ifdef BCMETH_MPSAFETX
1762 	} else {
1763 		/*
1764 		 * Either we are in a softintr thread already or some other
1765 		 * thread so just borrow it to do the send and save ourselves
1766 		 * the overhead of a fast soft int.
1767 		 */
1768 		bcmeth_soft_txintr(sc);
1769 	}
1770 #endif
1771 }
1772 
1773 int
1774 bcmeth_intr(void *arg)
1775 {
1776 	struct bcmeth_softc * const sc = arg;
1777 	uint32_t soft_flags = 0;
1778 	uint32_t work_flags = 0;
1779 	int rv = 0;
1780 
1781 	mutex_enter(sc->sc_hwlock);
1782 
1783 	uint32_t intmask = sc->sc_intmask;
1784 	BCMETH_EVCNT_INCR(sc->sc_ev_intr);
1785 
1786 	for (;;) {
1787 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
1788 		intstatus &= intmask;
1789 		bcmeth_write_4(sc, GMAC_INTSTATUS, intstatus);	/* write 1 to clear */
1790 		if (intstatus == 0) {
1791 			break;
1792 		}
1793 #if 0
1794 		aprint_normal_dev(sc->sc_dev, "%s: intstatus=%#x intmask=%#x\n",
1795 		    __func__, intstatus, bcmeth_read_4(sc, GMAC_INTMASK));
1796 #endif
1797 		if (intstatus & RCVINT) {
1798 			struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
1799 			intmask &= ~RCVINT;
1800 
1801 			uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1802 			uint32_t descs = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1803 			if (descs < rxq->rxq_consumer - rxq->rxq_first) {
1804 				/*
1805 				 * We wrapped at the end so count how far
1806 				 * we are from the end.
1807 				 */
1808 				descs += rxq->rxq_last - rxq->rxq_consumer;
1809 			} else {
1810 				descs -= rxq->rxq_consumer - rxq->rxq_first;
1811 			}
1812 			/*
1813 			 * If we "timedout" we can't be hogging so use
1814 			 * softints.  If we exceeded then we might hogging
1815 			 * so let the workqueue deal with them.
1816 			 */
1817 			const uint32_t framecount = __SHIFTOUT(sc->sc_rcvlazy,
1818 			    INTRCVLAZY_FRAMECOUNT);
1819 			if (descs < framecount
1820 			    || (curcpu()->ci_curlwp->l_flag & LW_IDLE)) {
1821 				soft_flags |= SOFT_RXINTR;
1822 			} else {
1823 				work_flags |= WORK_RXINTR;
1824 			}
1825 		}
1826 
1827 		if (intstatus & XMTINT_0) {
1828 			intmask &= ~XMTINT_0;
1829 			soft_flags |= SOFT_TXINTR;
1830 		}
1831 
1832 		if (intstatus & RCVDESCUF) {
1833 			intmask &= ~RCVDESCUF;
1834 			work_flags |= WORK_RXUNDERFLOW;
1835 		}
1836 
1837 		intstatus &= intmask;
1838 		if (intstatus) {
1839 			aprint_error_dev(sc->sc_dev,
1840 			    "intr: intstatus=%#x\n", intstatus);
1841 			aprint_error_dev(sc->sc_dev,
1842 			    "rcvbase=%p/%#lx rcvptr=%#x rcvsts=%#x/%#x\n",
1843 			    sc->sc_rxq.rxq_first,
1844 			    sc->sc_rxq.rxq_descmap->dm_segs[0].ds_addr,
1845 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvptr),
1846 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts0),
1847 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts1));
1848 			aprint_error_dev(sc->sc_dev,
1849 			    "xmtbase=%p/%#lx xmtptr=%#x xmtsts=%#x/%#x\n",
1850 			    sc->sc_txq.txq_first,
1851 			    sc->sc_txq.txq_descmap->dm_segs[0].ds_addr,
1852 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtptr),
1853 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts0),
1854 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts1));
1855 			intmask &= ~intstatus;
1856 			work_flags |= WORK_REINIT;
1857 			break;
1858 		}
1859 	}
1860 
1861 	if (intmask != sc->sc_intmask) {
1862 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1863 	}
1864 
1865 	if (work_flags) {
1866 		if (sc->sc_work_flags == 0) {
1867 			workqueue_enqueue(sc->sc_workq, &sc->sc_work, NULL);
1868 		}
1869 		atomic_or_32(&sc->sc_work_flags, work_flags);
1870 		rv = 1;
1871 	}
1872 
1873 	if (soft_flags) {
1874 		if (sc->sc_soft_flags == 0) {
1875 			softint_schedule(sc->sc_soft_ih);
1876 		}
1877 		atomic_or_32(&sc->sc_soft_flags, soft_flags);
1878 		rv = 1;
1879 	}
1880 
1881 	mutex_exit(sc->sc_hwlock);
1882 
1883 	return rv;
1884 }
1885 
1886 #ifdef BCMETH_MPSAFETX
1887 void
1888 bcmeth_soft_txintr(struct bcmeth_softc *sc)
1889 {
1890 	mutex_enter(sc->sc_lock);
1891 	/*
1892 	 * Let's do what we came here for.  Consume transmitted
1893 	 * packets off the transmit ring.
1894 	 */
1895 	if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1896 	    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1897 		BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1898 	}
1899 	if (sc->sc_if.if_flags & IFF_RUNNING) {
1900 		mutex_spin_enter(sc->sc_hwlock);
1901 		sc->sc_intmask |= XMTINT_0;
1902 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1903 		mutex_spin_exit(sc->sc_hwlock);
1904 	}
1905 	mutex_exit(sc->sc_lock);
1906 }
1907 #endif /* BCMETH_MPSAFETX */
1908 
1909 void
1910 bcmeth_soft_intr(void *arg)
1911 {
1912 	struct bcmeth_softc * const sc = arg;
1913 	struct ifnet * const ifp = &sc->sc_if;
1914 	uint32_t intmask = 0;
1915 
1916 	mutex_enter(sc->sc_lock);
1917 
1918 	u_int soft_flags = atomic_swap_uint(&sc->sc_soft_flags, 0);
1919 
1920 	BCMETH_EVCNT_INCR(sc->sc_ev_soft_intr);
1921 
1922 	if ((soft_flags & SOFT_TXINTR)
1923 	    || bcmeth_txq_active_p(sc, &sc->sc_txq)) {
1924 		/*
1925 		 * Let's do what we came here for.  Consume transmitted
1926 		 * packets off the transmit ring.
1927 		 */
1928 		if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1929 		    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1930 			BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1931 		}
1932 		intmask |= XMTINT_0;
1933 	}
1934 
1935 	if (soft_flags & SOFT_RXINTR) {
1936 		/*
1937 		 * Let's consume
1938 		 */
1939 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
1940 		    sc->sc_rxq.rxq_threshold / 4)) {
1941 			/*
1942 			 * We've consumed a quarter of the ring and still have
1943 			 * more to do.  Refill the ring.
1944 			 */
1945 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
1946 		}
1947 		intmask |= RCVINT;
1948 	}
1949 
1950 	if (ifp->if_flags & IFF_RUNNING) {
1951 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
1952 		mutex_spin_enter(sc->sc_hwlock);
1953 		sc->sc_intmask |= intmask;
1954 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1955 		mutex_spin_exit(sc->sc_hwlock);
1956 	}
1957 
1958 	mutex_exit(sc->sc_lock);
1959 }
1960 
1961 void
1962 bcmeth_worker(struct work *wk, void *arg)
1963 {
1964 	struct bcmeth_softc * const sc = arg;
1965 	struct ifnet * const ifp = &sc->sc_if;
1966 	uint32_t intmask = 0;
1967 
1968 	mutex_enter(sc->sc_lock);
1969 
1970 	BCMETH_EVCNT_INCR(sc->sc_ev_work);
1971 
1972 	uint32_t work_flags = atomic_swap_32(&sc->sc_work_flags, 0);
1973 	if (work_flags & WORK_REINIT) {
1974 		int s = splnet();
1975 		sc->sc_soft_flags = 0;
1976 		bcmeth_ifinit(ifp);
1977 		splx(s);
1978 		work_flags &= ~WORK_RXUNDERFLOW;
1979 	}
1980 
1981 	if (work_flags & WORK_RXUNDERFLOW) {
1982 		struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
1983 		size_t threshold = 5 * rxq->rxq_threshold / 4;
1984 		if (threshold >= rxq->rxq_last - rxq->rxq_first) {
1985 			threshold = rxq->rxq_last - rxq->rxq_first - 1;
1986 		} else {
1987 			intmask |= RCVDESCUF;
1988 		}
1989 		aprint_normal_dev(sc->sc_dev,
1990 		    "increasing receive buffers from %zu to %zu\n",
1991 		    rxq->rxq_threshold, threshold);
1992 		rxq->rxq_threshold = threshold;
1993 	}
1994 
1995 	if (work_flags & WORK_RXINTR) {
1996 		/*
1997 		 * Let's consume
1998 		 */
1999 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
2000 		    sc->sc_rxq.rxq_threshold / 4)) {
2001 			/*
2002 			 * We've consumed a quarter of the ring and still have
2003 			 * more to do.  Refill the ring.
2004 			 */
2005 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
2006 		}
2007 		intmask |= RCVINT;
2008 	}
2009 
2010 	if (ifp->if_flags & IFF_RUNNING) {
2011 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
2012 #if 0
2013 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
2014 		if (intstatus & RCVINT) {
2015 			bcmeth_write_4(sc, GMAC_INTSTATUS, RCVINT);
2016 			work_flags |= WORK_RXINTR;
2017 			continue;
2018 		}
2019 #endif
2020 		mutex_spin_enter(sc->sc_hwlock);
2021 		sc->sc_intmask |= intmask;
2022 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
2023 		mutex_spin_exit(sc->sc_hwlock);
2024 	}
2025 
2026 	mutex_exit(sc->sc_lock);
2027 }
2028