xref: /netbsd-src/sys/arch/arm/broadcom/bcm53xx_eth.c (revision 501cd18a74d52bfcca7d9e7e3b0d472bbc870558)
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 1.29 2016/12/15 09:28:02 ozaki-r 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 		return;
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 		return;
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 		return;
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 		return;
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 		return;
352 	}
353 
354 	sc->sc_soft_ih = softint_establish(SOFTINT_MPSAFE | SOFTINT_NET,
355 	    bcmeth_soft_intr, sc);
356 
357 	sc->sc_ih = intr_establish(loc->loc_intrs[0], IPL_VM, IST_LEVEL,
358 	    bcmeth_intr, sc);
359 
360 	if (sc->sc_ih == NULL) {
361 		aprint_error_dev(self, "failed to establish interrupt %d\n",
362 		     loc->loc_intrs[0]);
363 	} else {
364 		aprint_normal_dev(self, "interrupting on irq %d\n",
365 		     loc->loc_intrs[0]);
366 	}
367 
368 	aprint_normal_dev(sc->sc_dev, "Ethernet address %s\n",
369 	    ether_sprintf(sc->sc_enaddr));
370 
371 	/*
372 	 * Since each port in plugged into the switch/flow-accelerator,
373 	 * we hard code at Gige Full-Duplex with Flow Control enabled.
374 	 */
375 	int ifmedia = IFM_ETHER|IFM_1000_T|IFM_FDX;
376 	//ifmedia |= IFM_FLOW|IFM_ETH_TXPAUSE|IFM_ETH_RXPAUSE;
377 	ifmedia_init(&sc->sc_media, IFM_IMASK, bcmeth_mediachange,
378 	    bcmeth_mediastatus);
379 	ifmedia_add(&sc->sc_media, ifmedia, 0, NULL);
380 	ifmedia_set(&sc->sc_media, ifmedia);
381 
382 	ec->ec_capabilities = ETHERCAP_VLAN_MTU | ETHERCAP_JUMBO_MTU;
383 
384 	strlcpy(ifp->if_xname, xname, IFNAMSIZ);
385 	ifp->if_softc = sc;
386 	ifp->if_baudrate = IF_Mbps(1000);
387 	ifp->if_capabilities = 0;
388 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
389 #ifdef BCMETH_MPSAFE
390 	ifp->if_flags2 = IFF2_MPSAFE;
391 #endif
392 	ifp->if_ioctl = bcmeth_ifioctl;
393 	ifp->if_start = bcmeth_ifstart;
394 	ifp->if_watchdog = bcmeth_ifwatchdog;
395 	ifp->if_init = bcmeth_ifinit;
396 	ifp->if_stop = bcmeth_ifstop;
397 	IFQ_SET_READY(&ifp->if_snd);
398 
399 	bcmeth_ifstop(ifp, true);
400 
401 	/*
402 	 * Attach the interface.
403 	 */
404 	if_initialize(ifp);
405 	ether_ifattach(ifp, sc->sc_enaddr);
406 	if_register(ifp);
407 
408 #ifdef BCMETH_COUNTERS
409 	evcnt_attach_dynamic(&sc->sc_ev_intr, EVCNT_TYPE_INTR,
410 	    NULL, xname, "intr");
411 	evcnt_attach_dynamic(&sc->sc_ev_soft_intr, EVCNT_TYPE_INTR,
412 	    NULL, xname, "soft intr");
413 	evcnt_attach_dynamic(&sc->sc_ev_work, EVCNT_TYPE_MISC,
414 	    NULL, xname, "work items");
415 	evcnt_attach_dynamic(&sc->sc_ev_tx_stall, EVCNT_TYPE_MISC,
416 	    NULL, xname, "tx stalls");
417 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_lo, EVCNT_TYPE_MISC,
418 	    NULL, xname, "rx badmagic lo");
419 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_hi, EVCNT_TYPE_MISC,
420 	    NULL, xname, "rx badmagic hi");
421 #endif
422 }
423 
424 static int
425 bcmeth_mediachange(struct ifnet *ifp)
426 {
427 	//struct bcmeth_softc * const sc = ifp->if_softc;
428 	return 0;
429 }
430 
431 static void
432 bcmeth_mediastatus(struct ifnet *ifp, struct ifmediareq *ifm)
433 {
434 	//struct bcmeth_softc * const sc = ifp->if_softc;
435 
436 	ifm->ifm_status = IFM_AVALID | IFM_ACTIVE;
437 	ifm->ifm_active = IFM_ETHER | IFM_FDX | IFM_1000_T;
438 }
439 
440 static uint64_t
441 bcmeth_macaddr_create(const uint8_t *enaddr)
442 {
443 	return (enaddr[3] << 0)			// UNIMAC_MAC_0
444 	    |  (enaddr[2] << 8)			// UNIMAC_MAC_0
445 	    |  (enaddr[1] << 16)		// UNIMAC_MAC_0
446 	    |  ((uint64_t)enaddr[0] << 24)	// UNIMAC_MAC_0
447 	    |  ((uint64_t)enaddr[5] << 32)	// UNIMAC_MAC_1
448 	    |  ((uint64_t)enaddr[4] << 40);	// UNIMAC_MAC_1
449 }
450 
451 static int
452 bcmeth_ifinit(struct ifnet *ifp)
453 {
454 	struct bcmeth_softc * const sc = ifp->if_softc;
455 	int error = 0;
456 
457 	sc->sc_maxfrm = max(ifp->if_mtu + 32, MCLBYTES);
458 	if (ifp->if_mtu > ETHERMTU_JUMBO)
459 		return error;
460 
461 	KASSERT(ifp->if_flags & IFF_UP);
462 
463 	/*
464 	 * Stop the interface
465 	 */
466 	bcmeth_ifstop(ifp, 0);
467 
468 	/*
469 	 * Reserve enough space at the front so that we can insert a maxsized
470 	 * link header and a VLAN tag.  Also make sure we have enough room for
471 	 * the rcvsts field as well.
472 	 */
473 	KASSERT(ALIGN(max_linkhdr) == max_linkhdr);
474 	KASSERTMSG(max_linkhdr > sizeof(struct ether_header), "%u > %zu",
475 	    max_linkhdr, sizeof(struct ether_header));
476 	sc->sc_rcvoffset = max_linkhdr + 4 - sizeof(struct ether_header);
477 	if (sc->sc_rcvoffset <= 4)
478 		sc->sc_rcvoffset += 4;
479 	KASSERT((sc->sc_rcvoffset & 3) == 2);
480 	KASSERT(sc->sc_rcvoffset <= __SHIFTOUT(RCVCTL_RCVOFFSET, RCVCTL_RCVOFFSET));
481 	KASSERT(sc->sc_rcvoffset >= 6);
482 
483 	/*
484 	 * If our frame size has changed (or it's our first time through)
485 	 * destroy the existing transmit mapcache.
486 	 */
487 	if (sc->sc_tx_mapcache != NULL
488 	    && sc->sc_maxfrm != sc->sc_tx_mapcache->dmc_maxmapsize) {
489 		bcmeth_mapcache_destroy(sc, sc->sc_tx_mapcache);
490 		sc->sc_tx_mapcache = NULL;
491 	}
492 
493 	if (sc->sc_tx_mapcache == NULL) {
494 		error = bcmeth_mapcache_create(sc, &sc->sc_tx_mapcache,
495 		    BCMETH_MAXTXMBUFS, sc->sc_maxfrm, BCMETH_NTXSEGS);
496 		if (error)
497 			return error;
498 	}
499 
500 	sc->sc_cmdcfg = NO_LENGTH_CHECK | PAUSE_IGNORE
501 	    | __SHIFTIN(ETH_SPEED_1000, ETH_SPEED)
502 	    | RX_ENA | TX_ENA;
503 
504 	if (ifp->if_flags & IFF_PROMISC) {
505 		sc->sc_cmdcfg |= PROMISC_EN;
506 	} else {
507 		sc->sc_cmdcfg &= ~PROMISC_EN;
508 	}
509 
510 	const uint8_t * const lladdr = CLLADDR(ifp->if_sadl);
511 	const uint64_t macstnaddr = bcmeth_macaddr_create(lladdr);
512 
513 	/*
514 	 * We make sure that a received Ethernet packet start on a non-word
515 	 * boundary so that the packet payload will be on a word boundary.
516 	 * So to check the destination address we keep around two words to
517 	 * quickly compare with.
518 	 */
519 #if __ARMEL__
520 	sc->sc_macaddr[0] = lladdr[0] | (lladdr[1] << 8);
521 	sc->sc_macaddr[1] = lladdr[2] | (lladdr[3] << 8)
522 	    | (lladdr[4] << 16) | (lladdr[5] << 24);
523 #else
524 	sc->sc_macaddr[0] = lladdr[1] | (lladdr[0] << 8);
525 	sc->sc_macaddr[1] = lladdr[5] | (lladdr[4] << 8)
526 	    | (lladdr[1] << 16) | (lladdr[2] << 24);
527 #endif
528 
529 	sc->sc_intmask = DESCPROTOERR|DATAERR|DESCERR;
530 
531 	/* 5. Load RCVADDR_LO with new pointer */
532 	bcmeth_rxq_reset(sc, &sc->sc_rxq);
533 
534 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
535 	    __SHIFTIN(sc->sc_rcvoffset, RCVCTL_RCVOFFSET)
536 	    | RCVCTL_PARITY_DIS
537 	    | RCVCTL_OFLOW_CONTINUE
538 	    | __SHIFTIN(3, RCVCTL_BURSTLEN));
539 
540 	/* 6. Load XMTADDR_LO with new pointer */
541 	bcmeth_txq_reset(sc, &sc->sc_txq);
542 
543 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl, XMTCTL_DMA_ACT_INDEX
544 	    | XMTCTL_PARITY_DIS
545 	    | __SHIFTIN(3, XMTCTL_BURSTLEN));
546 
547 	/* 7. Setup other UNIMAC registers */
548 	bcmeth_write_4(sc, UNIMAC_FRAME_LEN, sc->sc_maxfrm);
549 	bcmeth_write_4(sc, UNIMAC_MAC_0, (uint32_t)(macstnaddr >>  0));
550 	bcmeth_write_4(sc, UNIMAC_MAC_1, (uint32_t)(macstnaddr >> 32));
551 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, sc->sc_cmdcfg);
552 
553 	uint32_t devctl = bcmeth_read_4(sc, GMAC_DEVCONTROL);
554 	devctl |= RGMII_LINK_STATUS_SEL | NWAY_AUTO_POLL_EN | TXARB_STRICT_MODE;
555 	devctl &= ~FLOW_CTRL_MODE;
556 	devctl &= ~MIB_RD_RESET_EN;
557 	devctl &= ~RXQ_OVERFLOW_CTRL_SEL;
558 	devctl &= ~CPU_FLOW_CTRL_ON;
559 	bcmeth_write_4(sc, GMAC_DEVCONTROL, devctl);
560 
561 	/* Setup lazy receive (at most 1ms). */
562 	const struct cpu_softc * const cpu = curcpu()->ci_softc;
563 	sc->sc_rcvlazy =  __SHIFTIN(4, INTRCVLAZY_FRAMECOUNT)
564 	     | __SHIFTIN(cpu->cpu_clk.clk_apb / 1000, INTRCVLAZY_TIMEOUT);
565 	bcmeth_write_4(sc, GMAC_INTRCVLAZY, sc->sc_rcvlazy);
566 
567 	/* 11. Enable transmit queues in TQUEUE, and ensure that the transmit scheduling mode is correctly set in TCTRL. */
568 	sc->sc_intmask |= XMTINT_0|XMTUF;
569 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl,
570 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl) | XMTCTL_ENABLE);
571 
572 
573 	/* 12. Enable receive queues in RQUEUE, */
574 	sc->sc_intmask |= RCVINT|RCVDESCUF|RCVFIFOOF;
575 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
576 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl) | RCVCTL_ENABLE);
577 
578 	bcmeth_rxq_produce(sc, &sc->sc_rxq);	/* fill with rx buffers */
579 
580 #if 0
581 	aprint_normal_dev(sc->sc_dev,
582 	    "devctl=%#x ucmdcfg=%#x xmtctl=%#x rcvctl=%#x\n",
583 	    devctl, sc->sc_cmdcfg,
584 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl),
585 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl));
586 #endif
587 
588 	sc->sc_soft_flags = 0;
589 
590 	bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
591 
592 	ifp->if_flags |= IFF_RUNNING;
593 
594 	return error;
595 }
596 
597 static void
598 bcmeth_ifstop(struct ifnet *ifp, int disable)
599 {
600 	struct bcmeth_softc * const sc = ifp->if_softc;
601 	struct bcmeth_txqueue * const txq = &sc->sc_txq;
602 	struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
603 
604 	KASSERT(!cpu_intr_p());
605 
606 	sc->sc_soft_flags = 0;
607 	sc->sc_work_flags = 0;
608 
609 	/* Disable Rx processing */
610 	bcmeth_write_4(sc, rxq->rxq_reg_rcvctl,
611 	    bcmeth_read_4(sc, rxq->rxq_reg_rcvctl) & ~RCVCTL_ENABLE);
612 
613 	/* Disable Tx processing */
614 	bcmeth_write_4(sc, txq->txq_reg_xmtctl,
615 	    bcmeth_read_4(sc, txq->txq_reg_xmtctl) & ~XMTCTL_ENABLE);
616 
617 	/* Disable all interrupts */
618 	bcmeth_write_4(sc, GMAC_INTMASK, 0);
619 
620 	for (;;) {
621 		uint32_t tx0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
622 		uint32_t rx0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
623 		if (__SHIFTOUT(tx0, XMTSTATE) == XMTSTATE_DIS
624 		    && __SHIFTOUT(rx0, RCVSTATE) == RCVSTATE_DIS)
625 			break;
626 		delay(50);
627 	}
628 	/*
629 	 * Now reset the controller.
630 	 *
631 	 * 3. Set SW_RESET bit in UNIMAC_COMMAND_CONFIG register
632 	 * 4. Clear SW_RESET bit in UNIMAC_COMMAND_CONFIG register
633 	 */
634 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, SW_RESET);
635 	bcmeth_write_4(sc, GMAC_INTSTATUS, ~0);
636 	sc->sc_intmask = 0;
637 	ifp->if_flags &= ~IFF_RUNNING;
638 
639 	/*
640 	 * Let's consume any remaining transmitted packets.  And if we are
641 	 * disabling the interface, purge ourselves of any untransmitted
642 	 * packets.  But don't consume any received packets, just drop them.
643 	 * If we aren't disabling the interface, save the mbufs in the
644 	 * receive queue for reuse.
645 	 */
646 	bcmeth_rxq_purge(sc, &sc->sc_rxq, disable);
647 	bcmeth_txq_consume(sc, &sc->sc_txq);
648 	if (disable) {
649 		bcmeth_txq_purge(sc, &sc->sc_txq);
650 		IF_PURGE(&ifp->if_snd);
651 	}
652 
653 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, 0);
654 }
655 
656 static void
657 bcmeth_ifwatchdog(struct ifnet *ifp)
658 {
659 }
660 
661 static int
662 bcmeth_ifioctl(struct ifnet *ifp, u_long cmd, void *data)
663 {
664 	struct bcmeth_softc *sc  = ifp->if_softc;
665 	struct ifreq * const ifr = data;
666 	const int s = splnet();
667 	int error;
668 
669 	switch (cmd) {
670 	case SIOCSIFMEDIA:
671 	case SIOCGIFMEDIA:
672 		error = ifmedia_ioctl(ifp, ifr, &sc->sc_media, cmd);
673 		break;
674 
675 	default:
676 		error = ether_ioctl(ifp, cmd, data);
677 		if (error != ENETRESET)
678 			break;
679 
680 		if (cmd == SIOCADDMULTI || cmd == SIOCDELMULTI) {
681 			error = 0;
682 			break;
683 		}
684 		error = bcmeth_ifinit(ifp);
685 		break;
686 	}
687 
688 	splx(s);
689 	return error;
690 }
691 
692 static void
693 bcmeth_rxq_desc_presync(
694 	struct bcmeth_softc *sc,
695 	struct bcmeth_rxqueue *rxq,
696 	struct gmac_rxdb *rxdb,
697 	size_t count)
698 {
699 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
700 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
701 	    BUS_DMASYNC_PREWRITE);
702 }
703 
704 static void
705 bcmeth_rxq_desc_postsync(
706 	struct bcmeth_softc *sc,
707 	struct bcmeth_rxqueue *rxq,
708 	struct gmac_rxdb *rxdb,
709 	size_t count)
710 {
711 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
712 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
713 	    BUS_DMASYNC_POSTWRITE);
714 }
715 
716 static void
717 bcmeth_txq_desc_presync(
718 	struct bcmeth_softc *sc,
719 	struct bcmeth_txqueue *txq,
720 	struct gmac_txdb *txdb,
721 	size_t count)
722 {
723 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
724 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
725 	    BUS_DMASYNC_PREWRITE);
726 }
727 
728 static void
729 bcmeth_txq_desc_postsync(
730 	struct bcmeth_softc *sc,
731 	struct bcmeth_txqueue *txq,
732 	struct gmac_txdb *txdb,
733 	size_t count)
734 {
735 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
736 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
737 	    BUS_DMASYNC_POSTWRITE);
738 }
739 
740 static bus_dmamap_t
741 bcmeth_mapcache_get(
742 	struct bcmeth_softc *sc,
743 	struct bcmeth_mapcache *dmc)
744 {
745 	KASSERT(dmc->dmc_nmaps > 0);
746 	KASSERT(dmc->dmc_maps[dmc->dmc_nmaps-1] != NULL);
747 	return dmc->dmc_maps[--dmc->dmc_nmaps];
748 }
749 
750 static void
751 bcmeth_mapcache_put(
752 	struct bcmeth_softc *sc,
753 	struct bcmeth_mapcache *dmc,
754 	bus_dmamap_t map)
755 {
756 	KASSERT(map != NULL);
757 	KASSERT(dmc->dmc_nmaps < dmc->dmc_maxmaps);
758 	dmc->dmc_maps[dmc->dmc_nmaps++] = map;
759 }
760 
761 static void
762 bcmeth_mapcache_destroy(
763 	struct bcmeth_softc *sc,
764 	struct bcmeth_mapcache *dmc)
765 {
766 	const size_t dmc_size =
767 	    offsetof(struct bcmeth_mapcache, dmc_maps[dmc->dmc_maxmaps]);
768 
769 	for (u_int i = 0; i < dmc->dmc_maxmaps; i++) {
770 		bus_dmamap_destroy(sc->sc_dmat, dmc->dmc_maps[i]);
771 	}
772 	kmem_intr_free(dmc, dmc_size);
773 }
774 
775 static int
776 bcmeth_mapcache_create(
777 	struct bcmeth_softc *sc,
778 	struct bcmeth_mapcache **dmc_p,
779 	size_t maxmaps,
780 	size_t maxmapsize,
781 	size_t maxseg)
782 {
783 	const size_t dmc_size =
784 	    offsetof(struct bcmeth_mapcache, dmc_maps[maxmaps]);
785 	struct bcmeth_mapcache * const dmc =
786 		kmem_intr_zalloc(dmc_size, KM_NOSLEEP);
787 
788 	dmc->dmc_maxmaps = maxmaps;
789 	dmc->dmc_nmaps = maxmaps;
790 	dmc->dmc_maxmapsize = maxmapsize;
791 	dmc->dmc_maxseg = maxseg;
792 
793 	for (u_int i = 0; i < maxmaps; i++) {
794 		int error = bus_dmamap_create(sc->sc_dmat, dmc->dmc_maxmapsize,
795 		     dmc->dmc_maxseg, dmc->dmc_maxmapsize, 0,
796 		     BUS_DMA_WAITOK|BUS_DMA_ALLOCNOW, &dmc->dmc_maps[i]);
797 		if (error) {
798 			aprint_error_dev(sc->sc_dev,
799 			    "failed to creat dma map cache "
800 			    "entry %u of %zu: %d\n",
801 			    i, maxmaps, error);
802 			while (i-- > 0) {
803 				bus_dmamap_destroy(sc->sc_dmat,
804 				    dmc->dmc_maps[i]);
805 			}
806 			kmem_intr_free(dmc, dmc_size);
807 			return error;
808 		}
809 		KASSERT(dmc->dmc_maps[i] != NULL);
810 	}
811 
812 	*dmc_p = dmc;
813 
814 	return 0;
815 }
816 
817 #if 0
818 static void
819 bcmeth_dmamem_free(
820 	bus_dma_tag_t dmat,
821 	size_t map_size,
822 	bus_dma_segment_t *seg,
823 	bus_dmamap_t map,
824 	void *kvap)
825 {
826 	bus_dmamap_destroy(dmat, map);
827 	bus_dmamem_unmap(dmat, kvap, map_size);
828 	bus_dmamem_free(dmat, seg, 1);
829 }
830 #endif
831 
832 static int
833 bcmeth_dmamem_alloc(
834 	bus_dma_tag_t dmat,
835 	size_t map_size,
836 	bus_dma_segment_t *seg,
837 	bus_dmamap_t *map,
838 	void **kvap)
839 {
840 	int error;
841 	int nseg;
842 
843 	*kvap = NULL;
844 	*map = NULL;
845 
846 	error = bus_dmamem_alloc(dmat, map_size, 2*PAGE_SIZE, 0,
847 	   seg, 1, &nseg, 0);
848 	if (error)
849 		return error;
850 
851 	KASSERT(nseg == 1);
852 
853 	error = bus_dmamem_map(dmat, seg, nseg, map_size, (void **)kvap, 0);
854 	if (error == 0) {
855 		error = bus_dmamap_create(dmat, map_size, 1, map_size, 0, 0,
856 		    map);
857 		if (error == 0) {
858 			error = bus_dmamap_load(dmat, *map, *kvap, map_size,
859 			    NULL, 0);
860 			if (error == 0)
861 				return 0;
862 			bus_dmamap_destroy(dmat, *map);
863 			*map = NULL;
864 		}
865 		bus_dmamem_unmap(dmat, *kvap, map_size);
866 		*kvap = NULL;
867 	}
868 	bus_dmamem_free(dmat, seg, nseg);
869 	return 0;
870 }
871 
872 static struct mbuf *
873 bcmeth_rx_buf_alloc(
874 	struct bcmeth_softc *sc)
875 {
876 	struct mbuf *m = m_gethdr(M_DONTWAIT, MT_DATA);
877 	if (m == NULL) {
878 		printf("%s:%d: %s\n", __func__, __LINE__, "m_gethdr");
879 		return NULL;
880 	}
881 	MCLGET(m, M_DONTWAIT);
882 	if ((m->m_flags & M_EXT) == 0) {
883 		printf("%s:%d: %s\n", __func__, __LINE__, "MCLGET");
884 		m_freem(m);
885 		return NULL;
886 	}
887 	m->m_len = m->m_pkthdr.len = m->m_ext.ext_size;
888 
889 	bus_dmamap_t map = bcmeth_mapcache_get(sc, sc->sc_rx_mapcache);
890 	if (map == NULL) {
891 		printf("%s:%d: %s\n", __func__, __LINE__, "map get");
892 		m_freem(m);
893 		return NULL;
894 	}
895 	M_SETCTX(m, map);
896 	m->m_len = m->m_pkthdr.len = MCLBYTES;
897 	int error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
898 	    BUS_DMA_READ|BUS_DMA_NOWAIT);
899 	if (error) {
900 		aprint_error_dev(sc->sc_dev, "fail to load rx dmamap: %d\n",
901 		    error);
902 		M_SETCTX(m, NULL);
903 		m_freem(m);
904 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
905 		return NULL;
906 	}
907 	KASSERT(map->dm_mapsize == MCLBYTES);
908 #ifdef BCMETH_RCVMAGIC
909 	*mtod(m, uint32_t *) = htole32(BCMETH_RCVMAGIC);
910 	bus_dmamap_sync(sc->sc_dmat, map, 0, sizeof(uint32_t),
911 	    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
912 	bus_dmamap_sync(sc->sc_dmat, map, sizeof(uint32_t),
913 	    map->dm_mapsize - sizeof(uint32_t), BUS_DMASYNC_PREREAD);
914 #else
915 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
916 	    BUS_DMASYNC_PREREAD);
917 #endif
918 
919 	return m;
920 }
921 
922 static void
923 bcmeth_rx_map_unload(
924 	struct bcmeth_softc *sc,
925 	struct mbuf *m)
926 {
927 	KASSERT(m);
928 	for (; m != NULL; m = m->m_next) {
929 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
930 		KASSERT(map);
931 		KASSERT(map->dm_mapsize == MCLBYTES);
932 		bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_len,
933 		    BUS_DMASYNC_POSTREAD);
934 		bus_dmamap_unload(sc->sc_dmat, map);
935 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
936 		M_SETCTX(m, NULL);
937 	}
938 }
939 
940 static bool
941 bcmeth_rxq_produce(
942 	struct bcmeth_softc *sc,
943 	struct bcmeth_rxqueue *rxq)
944 {
945 	struct gmac_rxdb *producer = rxq->rxq_producer;
946 	bool produced = false;
947 
948 	while (rxq->rxq_inuse < rxq->rxq_threshold) {
949 		struct mbuf *m;
950 		IF_DEQUEUE(&sc->sc_rx_bufcache, m);
951 		if (m == NULL) {
952 			m = bcmeth_rx_buf_alloc(sc);
953 			if (m == NULL) {
954 				printf("%s: bcmeth_rx_buf_alloc failed\n", __func__);
955 				break;
956 			}
957 		}
958 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
959 		KASSERT(map);
960 
961 		producer->rxdb_buflen = htole32(MCLBYTES);
962 		producer->rxdb_addrlo = htole32(map->dm_segs[0].ds_addr);
963 		producer->rxdb_flags &= htole32(RXDB_FLAG_ET);
964 		*rxq->rxq_mtail = m;
965 		rxq->rxq_mtail = &m->m_next;
966 		m->m_len = MCLBYTES;
967 		m->m_next = NULL;
968 		rxq->rxq_inuse++;
969 		if (++producer == rxq->rxq_last) {
970 			membar_producer();
971 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
972 			    rxq->rxq_last - rxq->rxq_producer);
973 			producer = rxq->rxq_producer = rxq->rxq_first;
974 		}
975 		produced = true;
976 	}
977 	if (produced) {
978 		membar_producer();
979 		if (producer != rxq->rxq_producer) {
980 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
981 			    producer - rxq->rxq_producer);
982 			rxq->rxq_producer = producer;
983 		}
984 		bcmeth_write_4(sc, rxq->rxq_reg_rcvptr,
985 		    rxq->rxq_descmap->dm_segs[0].ds_addr
986 		    + ((uintptr_t)producer & RCVPTR));
987 	}
988 	return true;
989 }
990 
991 static void
992 bcmeth_rx_input(
993 	struct bcmeth_softc *sc,
994 	struct mbuf *m,
995 	uint32_t rxdb_flags)
996 {
997 	struct ifnet * const ifp = &sc->sc_if;
998 
999 	bcmeth_rx_map_unload(sc, m);
1000 
1001 	m_adj(m, sc->sc_rcvoffset);
1002 
1003 	/*
1004 	 * If we are in promiscuous mode and this isn't a multicast, check the
1005 	 * destination address to make sure it matches our own.  If it doesn't,
1006 	 * mark the packet as being received promiscuously.
1007 	 */
1008 	if ((sc->sc_cmdcfg & PROMISC_EN)
1009 	    && (m->m_data[0] & 1) == 0
1010 	    && (*(uint16_t *)&m->m_data[0] != sc->sc_macaddr[0]
1011 		|| *(uint32_t *)&m->m_data[2] != sc->sc_macaddr[1])) {
1012 		m->m_flags |= M_PROMISC;
1013 	}
1014 	m_set_rcvif(m, ifp);
1015 
1016 	ifp->if_ibytes += m->m_pkthdr.len;
1017 
1018 	/*
1019 	 * Let's give it to the network subsystm to deal with.
1020 	 */
1021 #ifdef BCMETH_MPSAFE
1022 	mutex_exit(sc->sc_lock);
1023 	if_input(ifp, m);
1024 	mutex_enter(sc->sc_lock);
1025 #else
1026 	int s = splnet();
1027 	if_input(ifp, m);
1028 	splx(s);
1029 #endif
1030 }
1031 
1032 static bool
1033 bcmeth_rxq_consume(
1034 	struct bcmeth_softc *sc,
1035 	struct bcmeth_rxqueue *rxq,
1036 	size_t atmost)
1037 {
1038 	struct ifnet * const ifp = &sc->sc_if;
1039 	struct gmac_rxdb *consumer = rxq->rxq_consumer;
1040 	size_t rxconsumed = 0;
1041 	bool didconsume = false;
1042 
1043 	while (atmost-- > 0) {
1044 		if (consumer == rxq->rxq_producer) {
1045 			KASSERT(rxq->rxq_inuse == 0);
1046 			break;
1047 		}
1048 
1049 		uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1050 		uint32_t currdscr = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1051 		if (consumer == rxq->rxq_first + currdscr) {
1052 			break;
1053 		}
1054 		bcmeth_rxq_desc_postsync(sc, rxq, consumer, 1);
1055 
1056 		/*
1057 		 * We own this packet again.  Copy the rxsts word from it.
1058 		 */
1059 		rxconsumed++;
1060 		didconsume = true;
1061 		uint32_t rxsts;
1062 		KASSERT(rxq->rxq_mhead != NULL);
1063 		bus_dmamap_t map = M_GETCTX(rxq->rxq_mhead, bus_dmamap_t);
1064 		bus_dmamap_sync(sc->sc_dmat, map, 0, arm_dcache_align,
1065 		    BUS_DMASYNC_POSTREAD);
1066 		memcpy(&rxsts, rxq->rxq_mhead->m_data, 4);
1067 		rxsts = le32toh(rxsts);
1068 #if 0
1069 		KASSERTMSG(rxsts != BCMETH_RCVMAGIC, "currdscr=%u consumer=%zd",
1070 		    currdscr, consumer - rxq->rxq_first);
1071 #endif
1072 
1073 		/*
1074 		 * Get the count of descriptors.  Fetch the correct number
1075 		 * of mbufs.
1076 		 */
1077 #ifdef BCMETH_RCVMAGIC
1078 		size_t desc_count = rxsts != BCMETH_RCVMAGIC ? __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1 : 1;
1079 #else
1080 		size_t desc_count = __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1;
1081 #endif
1082 		struct mbuf *m = rxq->rxq_mhead;
1083 		struct mbuf *m_last = m;
1084 		for (size_t i = 1; i < desc_count; i++) {
1085 			if (++consumer == rxq->rxq_last) {
1086 				consumer = rxq->rxq_first;
1087 			}
1088 			KASSERTMSG(consumer != rxq->rxq_first + currdscr,
1089 			    "i=%zu rxsts=%#x desc_count=%zu currdscr=%u consumer=%zd",
1090 			    i, rxsts, desc_count, currdscr,
1091 			    consumer - rxq->rxq_first);
1092 			m_last = m_last->m_next;
1093 		}
1094 
1095 		/*
1096 		 * Now remove it/them from the list of enqueued mbufs.
1097 		 */
1098 		if ((rxq->rxq_mhead = m_last->m_next) == NULL)
1099 			rxq->rxq_mtail = &rxq->rxq_mhead;
1100 		m_last->m_next = NULL;
1101 
1102 #ifdef BCMETH_RCVMAGIC
1103 		if (rxsts == BCMETH_RCVMAGIC) {
1104 			ifp->if_ierrors++;
1105 			if ((m->m_ext.ext_paddr >> 28) == 8) {
1106 				BCMETH_EVCNT_INCR(sc->sc_ev_rx_badmagic_lo);
1107 			} else {
1108 				BCMETH_EVCNT_INCR( sc->sc_ev_rx_badmagic_hi);
1109 			}
1110 			IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1111 		} else
1112 #endif /* BCMETH_RCVMAGIC */
1113 		if (rxsts & (RXSTS_CRC_ERROR|RXSTS_OVERSIZED|RXSTS_PKT_OVERFLOW)) {
1114 			aprint_error_dev(sc->sc_dev, "[%zu]: count=%zu rxsts=%#x\n",
1115 			    consumer - rxq->rxq_first, desc_count, rxsts);
1116 			/*
1117 			 * We encountered an error, take the mbufs and add them
1118 			 * to the rx bufcache so we can quickly reuse them.
1119 			 */
1120 			ifp->if_ierrors++;
1121 			do {
1122 				struct mbuf *m0 = m->m_next;
1123 				m->m_next = NULL;
1124 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1125 				m = m0;
1126 			} while (m);
1127 		} else {
1128 			uint32_t framelen = __SHIFTOUT(rxsts, RXSTS_FRAMELEN);
1129 			framelen += sc->sc_rcvoffset;
1130 			m->m_pkthdr.len = framelen;
1131 			if (desc_count == 1) {
1132 				KASSERT(framelen <= MCLBYTES);
1133 				m->m_len = framelen;
1134 			} else {
1135 				m_last->m_len = framelen & (MCLBYTES - 1);
1136 			}
1137 
1138 #ifdef BCMETH_MPSAFE
1139 			/*
1140 			 * Wrap at the last entry!
1141 			 */
1142 			if (++consumer == rxq->rxq_last) {
1143 				KASSERT(consumer[-1].rxdb_flags & htole32(RXDB_FLAG_ET));
1144 				rxq->rxq_consumer = rxq->rxq_first;
1145 			} else {
1146 				rxq->rxq_consumer = consumer;
1147 			}
1148 			rxq->rxq_inuse -= rxconsumed;
1149 #endif /* BCMETH_MPSAFE */
1150 
1151 			/*
1152 			 * Receive the packet (which releases our lock)
1153 			 */
1154 			bcmeth_rx_input(sc, m, rxsts);
1155 
1156 #ifdef BCMETH_MPSAFE
1157 			/*
1158 			 * Since we had to give up our lock, we need to
1159 			 * refresh these.
1160 			 */
1161 			consumer = rxq->rxq_consumer;
1162 			rxconsumed = 0;
1163 			continue;
1164 #endif /* BCMETH_MPSAFE */
1165 		}
1166 
1167 		/*
1168 		 * Wrap at the last entry!
1169 		 */
1170 		if (++consumer == rxq->rxq_last) {
1171 			KASSERT(consumer[-1].rxdb_flags & htole32(RXDB_FLAG_ET));
1172 			consumer = rxq->rxq_first;
1173 		}
1174 	}
1175 
1176 	/*
1177 	 * Update queue info.
1178 	 */
1179 	rxq->rxq_consumer = consumer;
1180 	rxq->rxq_inuse -= rxconsumed;
1181 
1182 	/*
1183 	 * Did we consume anything?
1184 	 */
1185 	return didconsume;
1186 }
1187 
1188 static void
1189 bcmeth_rxq_purge(
1190 	struct bcmeth_softc *sc,
1191 	struct bcmeth_rxqueue *rxq,
1192 	bool discard)
1193 {
1194 	struct mbuf *m;
1195 
1196 	if ((m = rxq->rxq_mhead) != NULL) {
1197 		if (discard) {
1198 			bcmeth_rx_map_unload(sc, m);
1199 			m_freem(m);
1200 		} else {
1201 			while (m != NULL) {
1202 				struct mbuf *m0 = m->m_next;
1203 				m->m_next = NULL;
1204 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1205 				m = m0;
1206 			}
1207 		}
1208 
1209 	}
1210 
1211 	rxq->rxq_mhead = NULL;
1212 	rxq->rxq_mtail = &rxq->rxq_mhead;
1213 	rxq->rxq_inuse = 0;
1214 }
1215 
1216 static void
1217 bcmeth_rxq_reset(
1218 	struct bcmeth_softc *sc,
1219 	struct bcmeth_rxqueue *rxq)
1220 {
1221 	/*
1222 	 * sync all the descriptors
1223 	 */
1224 	bcmeth_rxq_desc_postsync(sc, rxq, rxq->rxq_first,
1225 	    rxq->rxq_last - rxq->rxq_first);
1226 
1227 	/*
1228 	 * Make sure we own all descriptors in the ring.
1229 	 */
1230 	struct gmac_rxdb *rxdb;
1231 	for (rxdb = rxq->rxq_first; rxdb < rxq->rxq_last - 1; rxdb++) {
1232 		rxdb->rxdb_flags = htole32(RXDB_FLAG_IC);
1233 	}
1234 
1235 	/*
1236 	 * Last descriptor has the wrap flag.
1237 	 */
1238 	rxdb->rxdb_flags = htole32(RXDB_FLAG_ET|RXDB_FLAG_IC);
1239 
1240 	/*
1241 	 * Reset the producer consumer indexes.
1242 	 */
1243 	rxq->rxq_consumer = rxq->rxq_first;
1244 	rxq->rxq_producer = rxq->rxq_first;
1245 	rxq->rxq_inuse = 0;
1246 	if (rxq->rxq_threshold < BCMETH_MINRXMBUFS)
1247 		rxq->rxq_threshold = BCMETH_MINRXMBUFS;
1248 
1249 	sc->sc_intmask |= RCVINT|RCVFIFOOF|RCVDESCUF;
1250 
1251 	/*
1252 	 * Restart the receiver at the first descriptor
1253 	 */
1254 	bcmeth_write_4(sc, rxq->rxq_reg_rcvaddrlo,
1255 	    rxq->rxq_descmap->dm_segs[0].ds_addr);
1256 }
1257 
1258 static int
1259 bcmeth_rxq_attach(
1260 	struct bcmeth_softc *sc,
1261 	struct bcmeth_rxqueue *rxq,
1262 	u_int qno)
1263 {
1264 	size_t desc_count = BCMETH_RINGSIZE / sizeof(rxq->rxq_first[0]);
1265 	int error;
1266 	void *descs;
1267 
1268 	KASSERT(desc_count == 256 || desc_count == 512);
1269 
1270 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1271 	   &rxq->rxq_descmap_seg, &rxq->rxq_descmap, &descs);
1272 	if (error)
1273 		return error;
1274 
1275 	memset(descs, 0, BCMETH_RINGSIZE);
1276 	rxq->rxq_first = descs;
1277 	rxq->rxq_last = rxq->rxq_first + desc_count;
1278 	rxq->rxq_consumer = descs;
1279 	rxq->rxq_producer = descs;
1280 
1281 	bcmeth_rxq_purge(sc, rxq, true);
1282 	bcmeth_rxq_reset(sc, rxq);
1283 
1284 	rxq->rxq_reg_rcvaddrlo = GMAC_RCVADDR_LOW;
1285 	rxq->rxq_reg_rcvctl = GMAC_RCVCONTROL;
1286 	rxq->rxq_reg_rcvptr = GMAC_RCVPTR;
1287 	rxq->rxq_reg_rcvsts0 = GMAC_RCVSTATUS0;
1288 	rxq->rxq_reg_rcvsts1 = GMAC_RCVSTATUS1;
1289 
1290 	return 0;
1291 }
1292 
1293 static bool
1294 bcmeth_txq_active_p(
1295 	struct bcmeth_softc * const sc,
1296 	struct bcmeth_txqueue *txq)
1297 {
1298 	return !IF_IS_EMPTY(&txq->txq_mbufs);
1299 }
1300 
1301 static bool
1302 bcmeth_txq_fillable_p(
1303 	struct bcmeth_softc * const sc,
1304 	struct bcmeth_txqueue *txq)
1305 {
1306 	return txq->txq_free >= txq->txq_threshold;
1307 }
1308 
1309 static int
1310 bcmeth_txq_attach(
1311 	struct bcmeth_softc *sc,
1312 	struct bcmeth_txqueue *txq,
1313 	u_int qno)
1314 {
1315 	size_t desc_count = BCMETH_RINGSIZE / sizeof(txq->txq_first[0]);
1316 	int error;
1317 	void *descs;
1318 
1319 	KASSERT(desc_count == 256 || desc_count == 512);
1320 
1321 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1322 	   &txq->txq_descmap_seg, &txq->txq_descmap, &descs);
1323 	if (error)
1324 		return error;
1325 
1326 	memset(descs, 0, BCMETH_RINGSIZE);
1327 	txq->txq_first = descs;
1328 	txq->txq_last = txq->txq_first + desc_count;
1329 	txq->txq_consumer = descs;
1330 	txq->txq_producer = descs;
1331 
1332 	IFQ_SET_MAXLEN(&txq->txq_mbufs, BCMETH_MAXTXMBUFS);
1333 
1334 	txq->txq_reg_xmtaddrlo = GMAC_XMTADDR_LOW;
1335 	txq->txq_reg_xmtctl = GMAC_XMTCONTROL;
1336 	txq->txq_reg_xmtptr = GMAC_XMTPTR;
1337 	txq->txq_reg_xmtsts0 = GMAC_XMTSTATUS0;
1338 	txq->txq_reg_xmtsts1 = GMAC_XMTSTATUS1;
1339 
1340 	bcmeth_txq_reset(sc, txq);
1341 
1342 	return 0;
1343 }
1344 
1345 static int
1346 bcmeth_txq_map_load(
1347 	struct bcmeth_softc *sc,
1348 	struct bcmeth_txqueue *txq,
1349 	struct mbuf *m)
1350 {
1351 	bus_dmamap_t map;
1352 	int error;
1353 
1354 	map = M_GETCTX(m, bus_dmamap_t);
1355 	if (map != NULL)
1356 		return 0;
1357 
1358 	map = bcmeth_mapcache_get(sc, sc->sc_tx_mapcache);
1359 	if (map == NULL)
1360 		return ENOMEM;
1361 
1362 	error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
1363 	    BUS_DMA_WRITE | BUS_DMA_NOWAIT);
1364 	if (error)
1365 		return error;
1366 
1367 	bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_pkthdr.len,
1368 	    BUS_DMASYNC_PREWRITE);
1369 	M_SETCTX(m, map);
1370 	return 0;
1371 }
1372 
1373 static void
1374 bcmeth_txq_map_unload(
1375 	struct bcmeth_softc *sc,
1376 	struct bcmeth_txqueue *txq,
1377 	struct mbuf *m)
1378 {
1379 	KASSERT(m);
1380 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1381 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1382 	    BUS_DMASYNC_POSTWRITE);
1383 	bus_dmamap_unload(sc->sc_dmat, map);
1384 	bcmeth_mapcache_put(sc, sc->sc_tx_mapcache, map);
1385 }
1386 
1387 static bool
1388 bcmeth_txq_produce(
1389 	struct bcmeth_softc *sc,
1390 	struct bcmeth_txqueue *txq,
1391 	struct mbuf *m)
1392 {
1393 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1394 
1395 	if (map->dm_nsegs > txq->txq_free)
1396 		return false;
1397 
1398 	/*
1399 	 * TCP Offload flag must be set in the first descriptor.
1400 	 */
1401 	struct gmac_txdb *producer = txq->txq_producer;
1402 	uint32_t first_flags = TXDB_FLAG_SF;
1403 	uint32_t last_flags = TXDB_FLAG_EF;
1404 
1405 	/*
1406 	 * If we've produced enough descriptors without consuming any
1407 	 * we need to ask for an interrupt to reclaim some.
1408 	 */
1409 	txq->txq_lastintr += map->dm_nsegs;
1410 	if (txq->txq_lastintr >= txq->txq_threshold
1411 	    || txq->txq_mbufs.ifq_len + 1 == txq->txq_mbufs.ifq_maxlen) {
1412 		txq->txq_lastintr = 0;
1413 		last_flags |= TXDB_FLAG_IC;
1414 	}
1415 
1416 	KASSERT(producer != txq->txq_last);
1417 
1418 	struct gmac_txdb *start = producer;
1419 	size_t count = map->dm_nsegs;
1420 	producer->txdb_flags |= htole32(first_flags);
1421 	producer->txdb_addrlo = htole32(map->dm_segs[0].ds_addr);
1422 	producer->txdb_buflen = htole32(map->dm_segs[0].ds_len);
1423 	for (u_int i = 1; i < map->dm_nsegs; i++) {
1424 #if 0
1425 		printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1426 		    le32toh(producer->txdb_flags),
1427 		    le32toh(producer->txdb_buflen),
1428 		    le32toh(producer->txdb_addrlo),
1429 		    le32toh(producer->txdb_addrhi));
1430 #endif
1431 		if (__predict_false(++producer == txq->txq_last)) {
1432 			bcmeth_txq_desc_presync(sc, txq, start,
1433 			    txq->txq_last - start);
1434 			count -= txq->txq_last - start;
1435 			producer = txq->txq_first;
1436 			start = txq->txq_first;
1437 		}
1438 		producer->txdb_addrlo = htole32(map->dm_segs[i].ds_addr);
1439 		producer->txdb_buflen = htole32(map->dm_segs[i].ds_len);
1440 	}
1441 	producer->txdb_flags |= htole32(last_flags);
1442 #if 0
1443 	printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1444 	    le32toh(producer->txdb_flags), le32toh(producer->txdb_buflen),
1445 	    le32toh(producer->txdb_addrlo), le32toh(producer->txdb_addrhi));
1446 #endif
1447 	if (count)
1448 		bcmeth_txq_desc_presync(sc, txq, start, count);
1449 
1450 	/*
1451 	 * Reduce free count by the number of segments we consumed.
1452 	 */
1453 	txq->txq_free -= map->dm_nsegs;
1454 	KASSERT(map->dm_nsegs == 1 || txq->txq_producer != producer);
1455 	KASSERT(map->dm_nsegs == 1 || (txq->txq_producer->txdb_flags & htole32(TXDB_FLAG_EF)) == 0);
1456 	KASSERT(producer->txdb_flags & htole32(TXDB_FLAG_EF));
1457 
1458 #if 0
1459 	printf("%s: mbuf %p: produced a %u byte packet in %u segments (%zd..%zd)\n",
1460 	    __func__, m, m->m_pkthdr.len, map->dm_nsegs,
1461 	    txq->txq_producer - txq->txq_first, producer - txq->txq_first);
1462 #endif
1463 
1464 	if (producer + 1 == txq->txq_last)
1465 		txq->txq_producer = txq->txq_first;
1466 	else
1467 		txq->txq_producer = producer + 1;
1468 	IF_ENQUEUE(&txq->txq_mbufs, m);
1469 
1470 	/*
1471 	 * Let the transmitter know there's more to do
1472 	 */
1473 	bcmeth_write_4(sc, txq->txq_reg_xmtptr,
1474 	    txq->txq_descmap->dm_segs[0].ds_addr
1475 	    + ((uintptr_t)txq->txq_producer & XMT_LASTDSCR));
1476 
1477 	return true;
1478 }
1479 
1480 static struct mbuf *
1481 bcmeth_copy_packet(struct mbuf *m)
1482 {
1483 	struct mbuf *mext = NULL;
1484 	size_t misalignment = 0;
1485 	size_t hlen = 0;
1486 
1487 	for (mext = m; mext != NULL; mext = mext->m_next) {
1488 		if (mext->m_flags & M_EXT) {
1489 			misalignment = mtod(mext, vaddr_t) & arm_dcache_align;
1490 			break;
1491 		}
1492 		hlen += m->m_len;
1493 	}
1494 
1495 	struct mbuf *n = m->m_next;
1496 	if (m != mext && hlen + misalignment <= MHLEN && false) {
1497 		KASSERT(m->m_pktdat <= m->m_data && m->m_data <= &m->m_pktdat[MHLEN - m->m_len]);
1498 		size_t oldoff = m->m_data - m->m_pktdat;
1499 		size_t off;
1500 		if (mext == NULL) {
1501 			off = (oldoff + hlen > MHLEN) ? 0 : oldoff;
1502 		} else {
1503 			off = MHLEN - (hlen + misalignment);
1504 		}
1505 		KASSERT(off + hlen + misalignment <= MHLEN);
1506 		if (((oldoff ^ off) & arm_dcache_align) != 0 || off < oldoff) {
1507 			memmove(&m->m_pktdat[off], m->m_data, m->m_len);
1508 			m->m_data = &m->m_pktdat[off];
1509 		}
1510 		m_copydata(n, 0, hlen - m->m_len, &m->m_data[m->m_len]);
1511 		m->m_len = hlen;
1512 		m->m_next = mext;
1513 		while (n != mext) {
1514 			n = m_free(n);
1515 		}
1516 		return m;
1517 	}
1518 
1519 	struct mbuf *m0 = m_gethdr(M_DONTWAIT, m->m_type);
1520 	if (m0 == NULL) {
1521 		return NULL;
1522 	}
1523 	M_COPY_PKTHDR(m0, m);
1524 	MCLAIM(m0, m->m_owner);
1525 	if (m0->m_pkthdr.len > MHLEN) {
1526 		MCLGET(m0, M_DONTWAIT);
1527 		if ((m0->m_flags & M_EXT) == 0) {
1528 			m_freem(m0);
1529 			return NULL;
1530 		}
1531 	}
1532 	m0->m_len = m->m_pkthdr.len;
1533 	m_copydata(m, 0, m0->m_len, mtod(m0, void *));
1534 	m_freem(m);
1535 	return m0;
1536 }
1537 
1538 static bool
1539 bcmeth_txq_enqueue(
1540 	struct bcmeth_softc *sc,
1541 	struct bcmeth_txqueue *txq)
1542 {
1543 	for (;;) {
1544 		if (IF_QFULL(&txq->txq_mbufs))
1545 			return false;
1546 		struct mbuf *m = txq->txq_next;
1547 		if (m == NULL) {
1548 			int s = splnet();
1549 			IF_DEQUEUE(&sc->sc_if.if_snd, m);
1550 			splx(s);
1551 			if (m == NULL)
1552 				return true;
1553 			M_SETCTX(m, NULL);
1554 		} else {
1555 			txq->txq_next = NULL;
1556 		}
1557 		/*
1558 		 * If LINK2 is set and this packet uses multiple mbufs,
1559 		 * consolidate it into a single mbuf.
1560 		 */
1561 		if (m->m_next != NULL && (sc->sc_if.if_flags & IFF_LINK2)) {
1562 			struct mbuf *m0 = bcmeth_copy_packet(m);
1563 			if (m0 == NULL) {
1564 				txq->txq_next = m;
1565 				return true;
1566 			}
1567 			m = m0;
1568 		}
1569 		int error = bcmeth_txq_map_load(sc, txq, m);
1570 		if (error) {
1571 			aprint_error_dev(sc->sc_dev,
1572 			    "discarded packet due to "
1573 			    "dmamap load failure: %d\n", error);
1574 			m_freem(m);
1575 			continue;
1576 		}
1577 		KASSERT(txq->txq_next == NULL);
1578 		if (!bcmeth_txq_produce(sc, txq, m)) {
1579 			txq->txq_next = m;
1580 			return false;
1581 		}
1582 		KASSERT(txq->txq_next == NULL);
1583 	}
1584 }
1585 
1586 static bool
1587 bcmeth_txq_consume(
1588 	struct bcmeth_softc *sc,
1589 	struct bcmeth_txqueue *txq)
1590 {
1591 	struct ifnet * const ifp = &sc->sc_if;
1592 	struct gmac_txdb *consumer = txq->txq_consumer;
1593 	size_t txfree = 0;
1594 
1595 #if 0
1596 	printf("%s: entry: free=%zu\n", __func__, txq->txq_free);
1597 #endif
1598 
1599 	for (;;) {
1600 		if (consumer == txq->txq_producer) {
1601 			txq->txq_consumer = consumer;
1602 			txq->txq_free += txfree;
1603 			txq->txq_lastintr -= min(txq->txq_lastintr, txfree);
1604 #if 0
1605 			printf("%s: empty: freed %zu descriptors going from %zu to %zu\n",
1606 			    __func__, txfree, txq->txq_free - txfree, txq->txq_free);
1607 #endif
1608 			KASSERT(txq->txq_lastintr == 0);
1609 			KASSERT(txq->txq_free == txq->txq_last - txq->txq_first - 1);
1610 			return true;
1611 		}
1612 		bcmeth_txq_desc_postsync(sc, txq, consumer, 1);
1613 		uint32_t s0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
1614 		if (consumer == txq->txq_first + __SHIFTOUT(s0, XMT_CURRDSCR)) {
1615 			txq->txq_consumer = consumer;
1616 			txq->txq_free += txfree;
1617 			txq->txq_lastintr -= min(txq->txq_lastintr, txfree);
1618 #if 0
1619 			printf("%s: freed %zu descriptors\n",
1620 			    __func__, txfree);
1621 #endif
1622 			return bcmeth_txq_fillable_p(sc, txq);
1623 		}
1624 
1625 		/*
1626 		 * If this is the last descriptor in the chain, get the
1627 		 * mbuf, free its dmamap, and free the mbuf chain itself.
1628 		 */
1629 		const uint32_t txdb_flags = le32toh(consumer->txdb_flags);
1630 		if (txdb_flags & TXDB_FLAG_EF) {
1631 			struct mbuf *m;
1632 
1633 			IF_DEQUEUE(&txq->txq_mbufs, m);
1634 			KASSERT(m);
1635 			bcmeth_txq_map_unload(sc, txq, m);
1636 #if 0
1637 			printf("%s: mbuf %p: consumed a %u byte packet\n",
1638 			    __func__, m, m->m_pkthdr.len);
1639 #endif
1640 			bpf_mtap(ifp, m);
1641 			ifp->if_opackets++;
1642 			ifp->if_obytes += m->m_pkthdr.len;
1643 			if (m->m_flags & M_MCAST)
1644 				ifp->if_omcasts++;
1645 			m_freem(m);
1646 		}
1647 
1648 		/*
1649 		 * We own this packet again.  Clear all flags except wrap.
1650 		 */
1651 		txfree++;
1652 
1653 		/*
1654 		 * Wrap at the last entry!
1655 		 */
1656 		if (txdb_flags & TXDB_FLAG_ET) {
1657 			consumer->txdb_flags = htole32(TXDB_FLAG_ET);
1658 			KASSERT(consumer + 1 == txq->txq_last);
1659 			consumer = txq->txq_first;
1660 		} else {
1661 			consumer->txdb_flags = 0;
1662 			consumer++;
1663 			KASSERT(consumer < txq->txq_last);
1664 		}
1665 	}
1666 }
1667 
1668 static void
1669 bcmeth_txq_purge(
1670 	struct bcmeth_softc *sc,
1671 	struct bcmeth_txqueue *txq)
1672 {
1673 	struct mbuf *m;
1674 	KASSERT((bcmeth_read_4(sc, UNIMAC_COMMAND_CONFIG) & TX_ENA) == 0);
1675 
1676 	for (;;) {
1677 		IF_DEQUEUE(&txq->txq_mbufs, m);
1678 		if (m == NULL)
1679 			break;
1680 		bcmeth_txq_map_unload(sc, txq, m);
1681 		m_freem(m);
1682 	}
1683 	if ((m = txq->txq_next) != NULL) {
1684 		txq->txq_next = NULL;
1685 		bcmeth_txq_map_unload(sc, txq, m);
1686 		m_freem(m);
1687 	}
1688 }
1689 
1690 static void
1691 bcmeth_txq_reset(
1692 	struct bcmeth_softc *sc,
1693 	struct bcmeth_txqueue *txq)
1694 {
1695 	/*
1696 	 * sync all the descriptors
1697 	 */
1698 	bcmeth_txq_desc_postsync(sc, txq, txq->txq_first,
1699 	    txq->txq_last - txq->txq_first);
1700 
1701 	/*
1702 	 * Make sure we own all descriptors in the ring.
1703 	 */
1704 	struct gmac_txdb *txdb;
1705 	for (txdb = txq->txq_first; txdb < txq->txq_last - 1; txdb++) {
1706 		txdb->txdb_flags = 0;
1707 	}
1708 
1709 	/*
1710 	 * Last descriptor has the wrap flag.
1711 	 */
1712 	txdb->txdb_flags = htole32(TXDB_FLAG_ET);
1713 
1714 	/*
1715 	 * Reset the producer consumer indexes.
1716 	 */
1717 	txq->txq_consumer = txq->txq_first;
1718 	txq->txq_producer = txq->txq_first;
1719 	txq->txq_free = txq->txq_last - txq->txq_first - 1;
1720 	txq->txq_threshold = txq->txq_free / 2;
1721 	txq->txq_lastintr = 0;
1722 
1723 	/*
1724 	 * What do we want to get interrupted on?
1725 	 */
1726 	sc->sc_intmask |= XMTINT_0 | XMTUF;
1727 
1728 	/*
1729 	 * Restart the transmiter at the first descriptor
1730 	 */
1731 	bcmeth_write_4(sc, txq->txq_reg_xmtaddrlo,
1732 	    txq->txq_descmap->dm_segs->ds_addr);
1733 }
1734 
1735 static void
1736 bcmeth_ifstart(struct ifnet *ifp)
1737 {
1738 	struct bcmeth_softc * const sc = ifp->if_softc;
1739 
1740 	if (__predict_false((ifp->if_flags & IFF_RUNNING) == 0)) {
1741 		return;
1742 	}
1743 
1744 #ifdef BCMETH_MPSAFETX
1745 	if (cpu_intr_p()) {
1746 #endif
1747 		atomic_or_uint(&sc->sc_soft_flags, SOFT_TXINTR);
1748 		softint_schedule(sc->sc_soft_ih);
1749 #ifdef BCMETH_MPSAFETX
1750 	} else {
1751 		/*
1752 		 * Either we are in a softintr thread already or some other
1753 		 * thread so just borrow it to do the send and save ourselves
1754 		 * the overhead of a fast soft int.
1755 		 */
1756 		bcmeth_soft_txintr(sc);
1757 	}
1758 #endif
1759 }
1760 
1761 int
1762 bcmeth_intr(void *arg)
1763 {
1764 	struct bcmeth_softc * const sc = arg;
1765 	uint32_t soft_flags = 0;
1766 	uint32_t work_flags = 0;
1767 	int rv = 0;
1768 
1769 	mutex_enter(sc->sc_hwlock);
1770 
1771 	uint32_t intmask = sc->sc_intmask;
1772 	BCMETH_EVCNT_INCR(sc->sc_ev_intr);
1773 
1774 	for (;;) {
1775 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
1776 		intstatus &= intmask;
1777 		bcmeth_write_4(sc, GMAC_INTSTATUS, intstatus);	/* write 1 to clear */
1778 		if (intstatus == 0) {
1779 			break;
1780 		}
1781 #if 0
1782 		aprint_normal_dev(sc->sc_dev, "%s: intstatus=%#x intmask=%#x\n",
1783 		    __func__, intstatus, bcmeth_read_4(sc, GMAC_INTMASK));
1784 #endif
1785 		if (intstatus & RCVINT) {
1786 			struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
1787 			intmask &= ~RCVINT;
1788 
1789 			uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1790 			uint32_t descs = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1791 			if (descs < rxq->rxq_consumer - rxq->rxq_first) {
1792 				/*
1793 				 * We wrapped at the end so count how far
1794 				 * we are from the end.
1795 				 */
1796 				descs += rxq->rxq_last - rxq->rxq_consumer;
1797 			} else {
1798 				descs -= rxq->rxq_consumer - rxq->rxq_first;
1799 			}
1800 			/*
1801 			 * If we "timedout" we can't be hogging so use
1802 			 * softints.  If we exceeded then we might hogging
1803 			 * so let the workqueue deal with them.
1804 			 */
1805 			const uint32_t framecount = __SHIFTOUT(sc->sc_rcvlazy, INTRCVLAZY_FRAMECOUNT);
1806 			if (descs < framecount
1807 			    || (curcpu()->ci_curlwp->l_flag & LW_IDLE)) {
1808 				soft_flags |= SOFT_RXINTR;
1809 			} else {
1810 				work_flags |= WORK_RXINTR;
1811 			}
1812 		}
1813 
1814 		if (intstatus & XMTINT_0) {
1815 			intmask &= ~XMTINT_0;
1816 			soft_flags |= SOFT_TXINTR;
1817 		}
1818 
1819 		if (intstatus & RCVDESCUF) {
1820 			intmask &= ~RCVDESCUF;
1821 			work_flags |= WORK_RXUNDERFLOW;
1822 		}
1823 
1824 		intstatus &= intmask;
1825 		if (intstatus) {
1826 			aprint_error_dev(sc->sc_dev,
1827 			    "intr: intstatus=%#x\n", intstatus);
1828 			aprint_error_dev(sc->sc_dev,
1829 			    "rcvbase=%p/%#lx rcvptr=%#x rcvsts=%#x/%#x\n",
1830 			    sc->sc_rxq.rxq_first,
1831 			    sc->sc_rxq.rxq_descmap->dm_segs[0].ds_addr,
1832 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvptr),
1833 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts0),
1834 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts1));
1835 			aprint_error_dev(sc->sc_dev,
1836 			    "xmtbase=%p/%#lx xmtptr=%#x xmtsts=%#x/%#x\n",
1837 			    sc->sc_txq.txq_first,
1838 			    sc->sc_txq.txq_descmap->dm_segs[0].ds_addr,
1839 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtptr),
1840 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts0),
1841 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts1));
1842 			intmask &= ~intstatus;
1843 			work_flags |= WORK_REINIT;
1844 			break;
1845 		}
1846 	}
1847 
1848 	if (intmask != sc->sc_intmask) {
1849 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1850 	}
1851 
1852 	if (work_flags) {
1853 		if (sc->sc_work_flags == 0) {
1854 			workqueue_enqueue(sc->sc_workq, &sc->sc_work, NULL);
1855 		}
1856 		atomic_or_32(&sc->sc_work_flags, work_flags);
1857 		rv = 1;
1858 	}
1859 
1860 	if (soft_flags) {
1861 		if (sc->sc_soft_flags == 0) {
1862 			softint_schedule(sc->sc_soft_ih);
1863 		}
1864 		atomic_or_32(&sc->sc_soft_flags, soft_flags);
1865 		rv = 1;
1866 	}
1867 
1868 	mutex_exit(sc->sc_hwlock);
1869 
1870 	return rv;
1871 }
1872 
1873 #ifdef BCMETH_MPSAFETX
1874 void
1875 bcmeth_soft_txintr(struct bcmeth_softc *sc)
1876 {
1877 	mutex_enter(sc->sc_lock);
1878 	/*
1879 	 * Let's do what we came here for.  Consume transmitted
1880 	 * packets off the the transmit ring.
1881 	 */
1882 	if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1883 	    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1884 		BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1885 		sc->sc_if.if_flags |= IFF_OACTIVE;
1886 	} else {
1887 		sc->sc_if.if_flags &= ~IFF_OACTIVE;
1888 	}
1889 	if (sc->sc_if.if_flags & IFF_RUNNING) {
1890 		mutex_spin_enter(sc->sc_hwlock);
1891 		sc->sc_intmask |= XMTINT_0;
1892 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1893 		mutex_spin_exit(sc->sc_hwlock);
1894 	}
1895 	mutex_exit(sc->sc_lock);
1896 }
1897 #endif /* BCMETH_MPSAFETX */
1898 
1899 void
1900 bcmeth_soft_intr(void *arg)
1901 {
1902 	struct bcmeth_softc * const sc = arg;
1903 	struct ifnet * const ifp = &sc->sc_if;
1904 	uint32_t intmask = 0;
1905 
1906 	mutex_enter(sc->sc_lock);
1907 
1908 	u_int soft_flags = atomic_swap_uint(&sc->sc_soft_flags, 0);
1909 
1910 	BCMETH_EVCNT_INCR(sc->sc_ev_soft_intr);
1911 
1912 	if ((soft_flags & SOFT_TXINTR)
1913 	    || bcmeth_txq_active_p(sc, &sc->sc_txq)) {
1914 		/*
1915 		 * Let's do what we came here for.  Consume transmitted
1916 		 * packets off the the transmit ring.
1917 		 */
1918 		if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1919 		    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1920 			BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1921 			ifp->if_flags |= IFF_OACTIVE;
1922 		} else {
1923 			ifp->if_flags &= ~IFF_OACTIVE;
1924 		}
1925 		intmask |= XMTINT_0;
1926 	}
1927 
1928 	if (soft_flags & SOFT_RXINTR) {
1929 		/*
1930 		 * Let's consume
1931 		 */
1932 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
1933 		    sc->sc_rxq.rxq_threshold / 4)) {
1934 			/*
1935 			 * We've consumed a quarter of the ring and still have
1936 			 * more to do.  Refill the ring.
1937 			 */
1938 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
1939 		}
1940 		intmask |= RCVINT;
1941 	}
1942 
1943 	if (ifp->if_flags & IFF_RUNNING) {
1944 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
1945 		mutex_spin_enter(sc->sc_hwlock);
1946 		sc->sc_intmask |= intmask;
1947 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1948 		mutex_spin_exit(sc->sc_hwlock);
1949 	}
1950 
1951 	mutex_exit(sc->sc_lock);
1952 }
1953 
1954 void
1955 bcmeth_worker(struct work *wk, void *arg)
1956 {
1957 	struct bcmeth_softc * const sc = arg;
1958 	struct ifnet * const ifp = &sc->sc_if;
1959 	uint32_t intmask = 0;
1960 
1961 	mutex_enter(sc->sc_lock);
1962 
1963 	BCMETH_EVCNT_INCR(sc->sc_ev_work);
1964 
1965 	uint32_t work_flags = atomic_swap_32(&sc->sc_work_flags, 0);
1966 	if (work_flags & WORK_REINIT) {
1967 		int s = splnet();
1968 		sc->sc_soft_flags = 0;
1969 		bcmeth_ifinit(ifp);
1970 		splx(s);
1971 		work_flags &= ~WORK_RXUNDERFLOW;
1972 	}
1973 
1974 	if (work_flags & WORK_RXUNDERFLOW) {
1975 		struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
1976 		size_t threshold = 5 * rxq->rxq_threshold / 4;
1977 		if (threshold >= rxq->rxq_last - rxq->rxq_first) {
1978 			threshold = rxq->rxq_last - rxq->rxq_first - 1;
1979 		} else {
1980 			intmask |= RCVDESCUF;
1981 		}
1982 		aprint_normal_dev(sc->sc_dev,
1983 		    "increasing receive buffers from %zu to %zu\n",
1984 		    rxq->rxq_threshold, threshold);
1985 		rxq->rxq_threshold = threshold;
1986 	}
1987 
1988 	if (work_flags & WORK_RXINTR) {
1989 		/*
1990 		 * Let's consume
1991 		 */
1992 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
1993 		    sc->sc_rxq.rxq_threshold / 4)) {
1994 			/*
1995 			 * We've consumed a quarter of the ring and still have
1996 			 * more to do.  Refill the ring.
1997 			 */
1998 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
1999 		}
2000 		intmask |= RCVINT;
2001 	}
2002 
2003 	if (ifp->if_flags & IFF_RUNNING) {
2004 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
2005 #if 0
2006 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
2007 		if (intstatus & RCVINT) {
2008 			bcmeth_write_4(sc, GMAC_INTSTATUS, RCVINT);
2009 			work_flags |= WORK_RXINTR;
2010 			continue;
2011 		}
2012 #endif
2013 		mutex_spin_enter(sc->sc_hwlock);
2014 		sc->sc_intmask |= intmask;
2015 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
2016 		mutex_spin_exit(sc->sc_hwlock);
2017 	}
2018 
2019 	mutex_exit(sc->sc_lock);
2020 }
2021