xref: /netbsd-src/sys/arch/arm/broadcom/bcm53xx_eth.c (revision 6a493d6bc668897c91594964a732d38505b70cbb)
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 1.25 2013/10/28 22:51:16 matt 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/broadcom/bcm53xx_reg.h>
63 #include <arm/broadcom/bcm53xx_var.h>
64 
65 //#define BCMETH_MPSAFE
66 
67 #ifdef BCMETH_COUNTERS
68 #define	BCMETH_EVCNT_ADD(a,b)	((void)((a).ev_count += (b)))
69 #else
70 #define	BCMETH_EVCNT_ADD(a,b)	do { } while (/*CONSTCOND*/0)
71 #endif
72 #define	BCMETH_EVCNT_INCR(a)	BCMETH_EVCNT_ADD((a), 1)
73 
74 #define	BCMETH_MAXTXMBUFS	128
75 #define	BCMETH_NTXSEGS		30
76 #define	BCMETH_MAXRXMBUFS	255
77 #define	BCMETH_MINRXMBUFS	64
78 #define	BCMETH_NRXSEGS		1
79 #define	BCMETH_RINGSIZE		PAGE_SIZE
80 
81 #if 1
82 #define	BCMETH_RCVMAGIC		0xfeedface
83 #endif
84 
85 static int bcmeth_ccb_match(device_t, cfdata_t, void *);
86 static void bcmeth_ccb_attach(device_t, device_t, void *);
87 
88 struct bcmeth_txqueue {
89 	bus_dmamap_t txq_descmap;
90 	struct gmac_txdb *txq_consumer;
91 	struct gmac_txdb *txq_producer;
92 	struct gmac_txdb *txq_first;
93 	struct gmac_txdb *txq_last;
94 	struct ifqueue txq_mbufs;
95 	struct mbuf *txq_next;
96 	size_t txq_free;
97 	size_t txq_threshold;
98 	size_t txq_lastintr;
99 	bus_size_t txq_reg_xmtaddrlo;
100 	bus_size_t txq_reg_xmtptr;
101 	bus_size_t txq_reg_xmtctl;
102 	bus_size_t txq_reg_xmtsts0;
103 	bus_size_t txq_reg_xmtsts1;
104 	bus_dma_segment_t txq_descmap_seg;
105 };
106 
107 struct bcmeth_rxqueue {
108 	bus_dmamap_t rxq_descmap;
109 	struct gmac_rxdb *rxq_consumer;
110 	struct gmac_rxdb *rxq_producer;
111 	struct gmac_rxdb *rxq_first;
112 	struct gmac_rxdb *rxq_last;
113 	struct mbuf *rxq_mhead;
114 	struct mbuf **rxq_mtail;
115 	struct mbuf *rxq_mconsumer;
116 	size_t rxq_inuse;
117 	size_t rxq_threshold;
118 	bus_size_t rxq_reg_rcvaddrlo;
119 	bus_size_t rxq_reg_rcvptr;
120 	bus_size_t rxq_reg_rcvctl;
121 	bus_size_t rxq_reg_rcvsts0;
122 	bus_size_t rxq_reg_rcvsts1;
123 	bus_dma_segment_t rxq_descmap_seg;
124 };
125 
126 struct bcmeth_mapcache {
127 	u_int dmc_nmaps;
128 	u_int dmc_maxseg;
129 	u_int dmc_maxmaps;
130 	u_int dmc_maxmapsize;
131 	bus_dmamap_t dmc_maps[0];
132 };
133 
134 struct bcmeth_softc {
135 	device_t sc_dev;
136 	bus_space_tag_t sc_bst;
137 	bus_space_handle_t sc_bsh;
138 	bus_dma_tag_t sc_dmat;
139 	kmutex_t *sc_lock;
140 	kmutex_t *sc_hwlock;
141 	struct ethercom sc_ec;
142 #define	sc_if		sc_ec.ec_if
143 	struct ifmedia sc_media;
144 	void *sc_soft_ih;
145 	void *sc_ih;
146 
147 	struct bcmeth_rxqueue sc_rxq;
148 	struct bcmeth_txqueue sc_txq;
149 
150 	size_t sc_rcvoffset;
151 	uint32_t sc_macaddr[2];
152 	uint32_t sc_maxfrm;
153 	uint32_t sc_cmdcfg;
154 	uint32_t sc_intmask;
155 	uint32_t sc_rcvlazy;
156 	volatile uint32_t sc_soft_flags;
157 #define	SOFT_RXINTR		0x01
158 #define	SOFT_TXINTR		0x02
159 
160 #ifdef BCMETH_COUNTERS
161 	struct evcnt sc_ev_intr;
162 	struct evcnt sc_ev_soft_intr;
163 	struct evcnt sc_ev_work;
164 	struct evcnt sc_ev_tx_stall;
165 	struct evcnt sc_ev_rx_badmagic_lo;
166 	struct evcnt sc_ev_rx_badmagic_hi;
167 #endif
168 
169 	struct ifqueue sc_rx_bufcache;
170 	struct bcmeth_mapcache *sc_rx_mapcache;
171 	struct bcmeth_mapcache *sc_tx_mapcache;
172 
173 	struct workqueue *sc_workq;
174 	struct work sc_work;
175 
176 	volatile uint32_t sc_work_flags;
177 #define	WORK_RXINTR		0x01
178 #define	WORK_RXUNDERFLOW	0x02
179 #define	WORK_REINIT		0x04
180 
181 	uint8_t sc_enaddr[ETHER_ADDR_LEN];
182 };
183 
184 static void bcmeth_ifstart(struct ifnet *);
185 static void bcmeth_ifwatchdog(struct ifnet *);
186 static int bcmeth_ifinit(struct ifnet *);
187 static void bcmeth_ifstop(struct ifnet *, int);
188 static int bcmeth_ifioctl(struct ifnet *, u_long, void *);
189 
190 static int bcmeth_mapcache_create(struct bcmeth_softc *,
191     struct bcmeth_mapcache **, size_t, size_t, size_t);
192 static void bcmeth_mapcache_destroy(struct bcmeth_softc *,
193     struct bcmeth_mapcache *);
194 static bus_dmamap_t bcmeth_mapcache_get(struct bcmeth_softc *,
195     struct bcmeth_mapcache *);
196 static void bcmeth_mapcache_put(struct bcmeth_softc *,
197     struct bcmeth_mapcache *, bus_dmamap_t);
198 
199 static int bcmeth_txq_attach(struct bcmeth_softc *,
200     struct bcmeth_txqueue *, u_int);
201 static void bcmeth_txq_purge(struct bcmeth_softc *,
202     struct bcmeth_txqueue *);
203 static void bcmeth_txq_reset(struct bcmeth_softc *,
204     struct bcmeth_txqueue *);
205 static bool bcmeth_txq_consume(struct bcmeth_softc *,
206     struct bcmeth_txqueue *);
207 static bool bcmeth_txq_produce(struct bcmeth_softc *,
208     struct bcmeth_txqueue *, struct mbuf *m);
209 static bool bcmeth_txq_active_p(struct bcmeth_softc *,
210     struct bcmeth_txqueue *);
211 
212 static int bcmeth_rxq_attach(struct bcmeth_softc *,
213     struct bcmeth_rxqueue *, u_int);
214 static bool bcmeth_rxq_produce(struct bcmeth_softc *,
215     struct bcmeth_rxqueue *);
216 static void bcmeth_rxq_purge(struct bcmeth_softc *,
217     struct bcmeth_rxqueue *, bool);
218 static void bcmeth_rxq_reset(struct bcmeth_softc *,
219     struct bcmeth_rxqueue *);
220 
221 static int bcmeth_intr(void *);
222 #ifdef BCMETH_MPSAFETX
223 static void bcmeth_soft_txintr(struct bcmeth_softc *);
224 #endif
225 static void bcmeth_soft_intr(void *);
226 static void bcmeth_worker(struct work *, void *);
227 
228 static int bcmeth_mediachange(struct ifnet *);
229 static void bcmeth_mediastatus(struct ifnet *, struct ifmediareq *);
230 
231 static inline uint32_t
232 bcmeth_read_4(struct bcmeth_softc *sc, bus_size_t o)
233 {
234 	return bus_space_read_4(sc->sc_bst, sc->sc_bsh, o);
235 }
236 
237 static inline void
238 bcmeth_write_4(struct bcmeth_softc *sc, bus_size_t o, uint32_t v)
239 {
240 	bus_space_write_4(sc->sc_bst, sc->sc_bsh, o, v);
241 }
242 
243 CFATTACH_DECL_NEW(bcmeth_ccb, sizeof(struct bcmeth_softc),
244 	bcmeth_ccb_match, bcmeth_ccb_attach, NULL, NULL);
245 
246 static int
247 bcmeth_ccb_match(device_t parent, cfdata_t cf, void *aux)
248 {
249 	struct bcmccb_attach_args * const ccbaa = aux;
250 	const struct bcm_locators * const loc = &ccbaa->ccbaa_loc;
251 
252 	if (strcmp(cf->cf_name, loc->loc_name))
253 		return 0;
254 
255 #ifdef DIAGNOSTIC
256 	const int port = cf->cf_loc[BCMCCBCF_PORT];
257 #endif
258 	KASSERT(port == BCMCCBCF_PORT_DEFAULT || port == loc->loc_port);
259 
260 	return 1;
261 }
262 
263 static void
264 bcmeth_ccb_attach(device_t parent, device_t self, void *aux)
265 {
266 	struct bcmeth_softc * const sc = device_private(self);
267 	struct ethercom * const ec = &sc->sc_ec;
268 	struct ifnet * const ifp = &ec->ec_if;
269 	struct bcmccb_attach_args * const ccbaa = aux;
270 	const struct bcm_locators * const loc = &ccbaa->ccbaa_loc;
271 	const char * const xname = device_xname(self);
272 	prop_dictionary_t dict = device_properties(self);
273 	int error;
274 
275 	sc->sc_bst = ccbaa->ccbaa_ccb_bst;
276 	sc->sc_dmat = ccbaa->ccbaa_dmat;
277 	bus_space_subregion(sc->sc_bst, ccbaa->ccbaa_ccb_bsh,
278 	    loc->loc_offset, loc->loc_size, &sc->sc_bsh);
279 
280 	/*
281 	 * We need to use the coherent dma tag for the GMAC.
282 	 */
283 	sc->sc_dmat = &bcm53xx_coherent_dma_tag;
284 #if _ARM32_NEED_BUS_DMA_BOUNCE
285 	if (device_cfdata(self)->cf_flags & 2) {
286 		sc->sc_dmat = &bcm53xx_bounce_dma_tag;
287 	}
288 #endif
289 
290 	prop_data_t eaprop = prop_dictionary_get(dict, "mac-address");
291         if (eaprop == NULL) {
292 		uint32_t mac0 = bcmeth_read_4(sc, UNIMAC_MAC_0);
293 		uint32_t mac1 = bcmeth_read_4(sc, UNIMAC_MAC_1);
294 		if ((mac0 == 0 && mac1 == 0) || (mac1 & 1)) {
295 			aprint_error(": mac-address property is missing\n");
296 			return;
297 		}
298 		sc->sc_enaddr[0] = (mac0 >> 0) & 0xff;
299 		sc->sc_enaddr[1] = (mac0 >> 8) & 0xff;
300 		sc->sc_enaddr[2] = (mac0 >> 16) & 0xff;
301 		sc->sc_enaddr[3] = (mac0 >> 24) & 0xff;
302 		sc->sc_enaddr[4] = (mac1 >> 0) & 0xff;
303 		sc->sc_enaddr[5] = (mac1 >> 8) & 0xff;
304 	} else {
305 		KASSERT(prop_object_type(eaprop) == PROP_TYPE_DATA);
306 		KASSERT(prop_data_size(eaprop) == ETHER_ADDR_LEN);
307 		memcpy(sc->sc_enaddr, prop_data_data_nocopy(eaprop),
308 		    ETHER_ADDR_LEN);
309 	}
310 	sc->sc_dev = self;
311 	sc->sc_lock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_SOFTNET);
312 	sc->sc_hwlock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_VM);
313 
314 	bcmeth_write_4(sc, GMAC_INTMASK, 0);	// disable interrupts
315 
316 	aprint_naive("\n");
317 	aprint_normal(": Gigabit Ethernet Controller\n");
318 
319 	error = bcmeth_rxq_attach(sc, &sc->sc_rxq, 0);
320 	if (error) {
321 		aprint_error(": failed to init rxq: %d\n", error);
322 		return;
323 	}
324 
325 	error = bcmeth_txq_attach(sc, &sc->sc_txq, 0);
326 	if (error) {
327 		aprint_error(": failed to init txq: %d\n", error);
328 		return;
329 	}
330 
331 	error = bcmeth_mapcache_create(sc, &sc->sc_rx_mapcache,
332 	    BCMETH_MAXRXMBUFS, MCLBYTES, BCMETH_NRXSEGS);
333 	if (error) {
334 		aprint_error(": failed to allocate rx dmamaps: %d\n", error);
335 		return;
336 	}
337 
338 	error = bcmeth_mapcache_create(sc, &sc->sc_tx_mapcache,
339 	    BCMETH_MAXTXMBUFS, MCLBYTES, BCMETH_NTXSEGS);
340 	if (error) {
341 		aprint_error(": failed to allocate tx dmamaps: %d\n", error);
342 		return;
343 	}
344 
345 	error = workqueue_create(&sc->sc_workq, xname, bcmeth_worker, sc,
346 	    (PRI_USER + MAXPRI_USER) / 2, IPL_NET, WQ_MPSAFE|WQ_PERCPU);
347 	if (error) {
348 		aprint_error(": failed to create workqueue: %d\n", error);
349 		return;
350 	}
351 
352 	sc->sc_soft_ih = softint_establish(SOFTINT_MPSAFE | SOFTINT_NET,
353 	    bcmeth_soft_intr, sc);
354 
355 	sc->sc_ih = intr_establish(loc->loc_intrs[0], IPL_VM, IST_LEVEL,
356 	    bcmeth_intr, sc);
357 
358 	if (sc->sc_ih == NULL) {
359 		aprint_error_dev(self, "failed to establish interrupt %d\n",
360 		     loc->loc_intrs[0]);
361 	} else {
362 		aprint_normal_dev(self, "interrupting on irq %d\n",
363 		     loc->loc_intrs[0]);
364 	}
365 
366 	aprint_normal_dev(sc->sc_dev, "Ethernet address %s\n",
367 	    ether_sprintf(sc->sc_enaddr));
368 
369 	/*
370 	 * Since each port in plugged into the switch/flow-accelerator,
371 	 * we hard code at Gige Full-Duplex with Flow Control enabled.
372 	 */
373 	int ifmedia = IFM_ETHER|IFM_1000_T|IFM_FDX;
374 	//ifmedia |= IFM_FLOW|IFM_ETH_TXPAUSE|IFM_ETH_RXPAUSE;
375 	ifmedia_init(&sc->sc_media, IFM_IMASK, bcmeth_mediachange,
376 	    bcmeth_mediastatus);
377 	ifmedia_add(&sc->sc_media, ifmedia, 0, NULL);
378 	ifmedia_set(&sc->sc_media, ifmedia);
379 
380 	ec->ec_capabilities = ETHERCAP_VLAN_MTU | ETHERCAP_JUMBO_MTU;
381 
382 	strlcpy(ifp->if_xname, xname, IFNAMSIZ);
383 	ifp->if_softc = sc;
384 	ifp->if_baudrate = IF_Mbps(1000);
385 	ifp->if_capabilities = 0;
386 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
387 #ifdef BCMETH_MPSAFE
388 	ifp->if_flags2 = IFF2_MPSAFE;
389 #endif
390 	ifp->if_ioctl = bcmeth_ifioctl;
391 	ifp->if_start = bcmeth_ifstart;
392 	ifp->if_watchdog = bcmeth_ifwatchdog;
393 	ifp->if_init = bcmeth_ifinit;
394 	ifp->if_stop = bcmeth_ifstop;
395 	IFQ_SET_READY(&ifp->if_snd);
396 
397 	bcmeth_ifstop(ifp, true);
398 
399 	/*
400 	 * Attach the interface.
401 	 */
402 	if_attach(ifp);
403 	ether_ifattach(ifp, sc->sc_enaddr);
404 
405 #ifdef BCMETH_COUNTERS
406 	evcnt_attach_dynamic(&sc->sc_ev_intr, EVCNT_TYPE_INTR,
407 	    NULL, xname, "intr");
408 	evcnt_attach_dynamic(&sc->sc_ev_soft_intr, EVCNT_TYPE_INTR,
409 	    NULL, xname, "soft intr");
410 	evcnt_attach_dynamic(&sc->sc_ev_work, EVCNT_TYPE_MISC,
411 	    NULL, xname, "work items");
412 	evcnt_attach_dynamic(&sc->sc_ev_tx_stall, EVCNT_TYPE_MISC,
413 	    NULL, xname, "tx stalls");
414 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_lo, EVCNT_TYPE_MISC,
415 	    NULL, xname, "rx badmagic lo");
416 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_hi, EVCNT_TYPE_MISC,
417 	    NULL, xname, "rx badmagic hi");
418 #endif
419 }
420 
421 static int
422 bcmeth_mediachange(struct ifnet *ifp)
423 {
424 	//struct bcmeth_softc * const sc = ifp->if_softc;
425 	return 0;
426 }
427 
428 static void
429 bcmeth_mediastatus(struct ifnet *ifp, struct ifmediareq *ifm)
430 {
431 	//struct bcmeth_softc * const sc = ifp->if_softc;
432 
433 	ifm->ifm_status = IFM_AVALID | IFM_ACTIVE;
434 	ifm->ifm_active = IFM_ETHER | IFM_FDX | IFM_1000_T;
435 }
436 
437 static uint64_t
438 bcmeth_macaddr_create(const uint8_t *enaddr)
439 {
440 	return (enaddr[3] << 0)			// UNIMAC_MAC_0
441 	    |  (enaddr[2] << 8)			// UNIMAC_MAC_0
442 	    |  (enaddr[1] << 16)		// UNIMAC_MAC_0
443 	    |  ((uint64_t)enaddr[0] << 24)	// UNIMAC_MAC_0
444 	    |  ((uint64_t)enaddr[5] << 32)	// UNIMAC_MAC_1
445 	    |  ((uint64_t)enaddr[4] << 40);	// UNIMAC_MAC_1
446 }
447 
448 static int
449 bcmeth_ifinit(struct ifnet *ifp)
450 {
451 	struct bcmeth_softc * const sc = ifp->if_softc;
452 	int error = 0;
453 
454 	sc->sc_maxfrm = max(ifp->if_mtu + 32, MCLBYTES);
455 	if (ifp->if_mtu > ETHERMTU_JUMBO)
456 		return error;
457 
458 	KASSERT(ifp->if_flags & IFF_UP);
459 
460 	/*
461 	 * Stop the interface
462 	 */
463 	bcmeth_ifstop(ifp, 0);
464 
465 	/*
466 	 * Reserve enough space at the front so that we can insert a maxsized
467 	 * link header and a VLAN tag.  Also make sure we have enough room for
468 	 * the rcvsts field as well.
469 	 */
470 	KASSERT(ALIGN(max_linkhdr) == max_linkhdr);
471 	KASSERTMSG(max_linkhdr > sizeof(struct ether_header), "%u > %zu",
472 	    max_linkhdr, sizeof(struct ether_header));
473 	sc->sc_rcvoffset = max_linkhdr + 4 - sizeof(struct ether_header);
474 	if (sc->sc_rcvoffset <= 4)
475 		sc->sc_rcvoffset += 4;
476 	KASSERT((sc->sc_rcvoffset & 3) == 2);
477 	KASSERT(sc->sc_rcvoffset <= __SHIFTOUT(RCVCTL_RCVOFFSET, RCVCTL_RCVOFFSET));
478 	KASSERT(sc->sc_rcvoffset >= 6);
479 
480 	/*
481 	 * If our frame size has changed (or it's our first time through)
482 	 * destroy the existing transmit mapcache.
483 	 */
484 	if (sc->sc_tx_mapcache != NULL
485 	    && sc->sc_maxfrm != sc->sc_tx_mapcache->dmc_maxmapsize) {
486 		bcmeth_mapcache_destroy(sc, sc->sc_tx_mapcache);
487 		sc->sc_tx_mapcache = NULL;
488 	}
489 
490 	if (sc->sc_tx_mapcache == NULL) {
491 		error = bcmeth_mapcache_create(sc, &sc->sc_tx_mapcache,
492 		    BCMETH_MAXTXMBUFS, sc->sc_maxfrm, BCMETH_NTXSEGS);
493 		if (error)
494 			return error;
495 	}
496 
497 	sc->sc_cmdcfg = NO_LENGTH_CHECK | PAUSE_IGNORE
498 	    | __SHIFTIN(ETH_SPEED_1000, ETH_SPEED)
499 	    | RX_ENA | TX_ENA;
500 
501 	if (ifp->if_flags & IFF_PROMISC) {
502 		sc->sc_cmdcfg |= PROMISC_EN;
503 	} else {
504 		sc->sc_cmdcfg &= ~PROMISC_EN;
505 	}
506 
507 	const uint8_t * const lladdr = CLLADDR(ifp->if_sadl);
508 	const uint64_t macstnaddr = bcmeth_macaddr_create(lladdr);
509 
510 	/*
511 	 * We make sure that a received Ethernet packet start on a non-word
512 	 * boundary so that the packet payload will be on a word boundary.
513 	 * So to check the destination address we keep around two words to
514 	 * quickly compare with.
515 	 */
516 #if __ARMEL__
517 	sc->sc_macaddr[0] = lladdr[0] | (lladdr[1] << 8);
518 	sc->sc_macaddr[1] = lladdr[2] | (lladdr[3] << 8)
519 	    | (lladdr[4] << 16) | (lladdr[5] << 24);
520 #else
521 	sc->sc_macaddr[0] = lladdr[1] | (lladdr[0] << 8);
522 	sc->sc_macaddr[1] = lladdr[5] | (lladdr[4] << 8)
523 	    | (lladdr[1] << 16) | (lladdr[2] << 24);
524 #endif
525 
526 	sc->sc_intmask = DESCPROTOERR|DATAERR|DESCERR;
527 
528 	/* 5. Load RCVADDR_LO with new pointer */
529 	bcmeth_rxq_reset(sc, &sc->sc_rxq);
530 
531 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
532 	    __SHIFTIN(sc->sc_rcvoffset, RCVCTL_RCVOFFSET)
533 	    | RCVCTL_PARITY_DIS
534 	    | RCVCTL_OFLOW_CONTINUE
535 	    | __SHIFTIN(3, RCVCTL_BURSTLEN));
536 
537 	/* 6. Load XMTADDR_LO with new pointer */
538 	bcmeth_txq_reset(sc, &sc->sc_txq);
539 
540 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl, XMTCTL_DMA_ACT_INDEX
541 	    | XMTCTL_PARITY_DIS
542 	    | __SHIFTIN(3, XMTCTL_BURSTLEN));
543 
544 	/* 7. Setup other UNIMAC registers */
545 	bcmeth_write_4(sc, UNIMAC_FRAME_LEN, sc->sc_maxfrm);
546 	bcmeth_write_4(sc, UNIMAC_MAC_0, (uint32_t)(macstnaddr >>  0));
547 	bcmeth_write_4(sc, UNIMAC_MAC_1, (uint32_t)(macstnaddr >> 32));
548 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, sc->sc_cmdcfg);
549 
550 	uint32_t devctl = bcmeth_read_4(sc, GMAC_DEVCONTROL);
551 	devctl |= RGMII_LINK_STATUS_SEL | NWAY_AUTO_POLL_EN | TXARB_STRICT_MODE;
552 	devctl &= ~FLOW_CTRL_MODE;
553 	devctl &= ~MIB_RD_RESET_EN;
554 	devctl &= ~RXQ_OVERFLOW_CTRL_SEL;
555 	devctl &= ~CPU_FLOW_CTRL_ON;
556 	bcmeth_write_4(sc, GMAC_DEVCONTROL, devctl);
557 
558 	/* Setup lazy receive (at most 1ms). */
559 	const struct cpu_softc * const cpu = curcpu()->ci_softc;
560 	sc->sc_rcvlazy =  __SHIFTIN(4, INTRCVLAZY_FRAMECOUNT)
561 	     | __SHIFTIN(cpu->cpu_clk.clk_apb / 1000, INTRCVLAZY_TIMEOUT);
562 	bcmeth_write_4(sc, GMAC_INTRCVLAZY, sc->sc_rcvlazy);
563 
564 	/* 11. Enable transmit queues in TQUEUE, and ensure that the transmit scheduling mode is correctly set in TCTRL. */
565 	sc->sc_intmask |= XMTINT_0|XMTUF;
566 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl,
567 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl) | XMTCTL_ENABLE);
568 
569 
570 	/* 12. Enable receive queues in RQUEUE, */
571 	sc->sc_intmask |= RCVINT|RCVDESCUF|RCVFIFOOF;
572 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
573 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl) | RCVCTL_ENABLE);
574 
575 	bcmeth_rxq_produce(sc, &sc->sc_rxq);	/* fill with rx buffers */
576 
577 #if 0
578 	aprint_normal_dev(sc->sc_dev,
579 	    "devctl=%#x ucmdcfg=%#x xmtctl=%#x rcvctl=%#x\n",
580 	    devctl, sc->sc_cmdcfg,
581 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl),
582 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl));
583 #endif
584 
585 	sc->sc_soft_flags = 0;
586 
587 	bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
588 
589 	ifp->if_flags |= IFF_RUNNING;
590 
591 	return error;
592 }
593 
594 static void
595 bcmeth_ifstop(struct ifnet *ifp, int disable)
596 {
597 	struct bcmeth_softc * const sc = ifp->if_softc;
598 	struct bcmeth_txqueue * const txq = &sc->sc_txq;
599 	struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
600 
601 	KASSERT(!cpu_intr_p());
602 
603 	sc->sc_soft_flags = 0;
604 	sc->sc_work_flags = 0;
605 
606 	/* Disable Rx processing */
607 	bcmeth_write_4(sc, rxq->rxq_reg_rcvctl,
608 	    bcmeth_read_4(sc, rxq->rxq_reg_rcvctl) & ~RCVCTL_ENABLE);
609 
610 	/* Disable Tx processing */
611 	bcmeth_write_4(sc, txq->txq_reg_xmtctl,
612 	    bcmeth_read_4(sc, txq->txq_reg_xmtctl) & ~XMTCTL_ENABLE);
613 
614 	/* Disable all interrupts */
615 	bcmeth_write_4(sc, GMAC_INTMASK, 0);
616 
617 	for (;;) {
618 		uint32_t tx0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
619 		uint32_t rx0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
620 		if (__SHIFTOUT(tx0, XMTSTATE) == XMTSTATE_DIS
621 		    && __SHIFTOUT(rx0, RCVSTATE) == RCVSTATE_DIS)
622 			break;
623 		delay(50);
624 	}
625 	/*
626 	 * Now reset the controller.
627 	 *
628 	 * 3. Set SW_RESET bit in UNIMAC_COMMAND_CONFIG register
629 	 * 4. Clear SW_RESET bit in UNIMAC_COMMAND_CONFIG register
630 	 */
631 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, SW_RESET);
632 	bcmeth_write_4(sc, GMAC_INTSTATUS, ~0);
633 	sc->sc_intmask = 0;
634 	ifp->if_flags &= ~IFF_RUNNING;
635 
636 	/*
637 	 * Let's consume any remaining transmitted packets.  And if we are
638 	 * disabling the interface, purge ourselves of any untransmitted
639 	 * packets.  But don't consume any received packets, just drop them.
640 	 * If we aren't disabling the interface, save the mbufs in the
641 	 * receive queue for reuse.
642 	 */
643 	bcmeth_rxq_purge(sc, &sc->sc_rxq, disable);
644 	bcmeth_txq_consume(sc, &sc->sc_txq);
645 	if (disable) {
646 		bcmeth_txq_purge(sc, &sc->sc_txq);
647 		IF_PURGE(&ifp->if_snd);
648 	}
649 
650 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, 0);
651 }
652 
653 static void
654 bcmeth_ifwatchdog(struct ifnet *ifp)
655 {
656 }
657 
658 static int
659 bcmeth_ifioctl(struct ifnet *ifp, u_long cmd, void *data)
660 {
661 	struct bcmeth_softc *sc  = ifp->if_softc;
662 	struct ifreq * const ifr = data;
663 	const int s = splnet();
664 	int error;
665 
666 	switch (cmd) {
667 	case SIOCSIFMEDIA:
668 	case SIOCGIFMEDIA:
669 		error = ifmedia_ioctl(ifp, ifr, &sc->sc_media, cmd);
670 		break;
671 
672 	default:
673 		error = ether_ioctl(ifp, cmd, data);
674 		if (error != ENETRESET)
675 			break;
676 
677 		if (cmd == SIOCADDMULTI || cmd == SIOCDELMULTI) {
678 			error = 0;
679 			break;
680 		}
681 		error = bcmeth_ifinit(ifp);
682 		break;
683 	}
684 
685 	splx(s);
686 	return error;
687 }
688 
689 static void
690 bcmeth_rxq_desc_presync(
691 	struct bcmeth_softc *sc,
692 	struct bcmeth_rxqueue *rxq,
693 	struct gmac_rxdb *rxdb,
694 	size_t count)
695 {
696 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
697 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
698 	    BUS_DMASYNC_PREWRITE);
699 }
700 
701 static void
702 bcmeth_rxq_desc_postsync(
703 	struct bcmeth_softc *sc,
704 	struct bcmeth_rxqueue *rxq,
705 	struct gmac_rxdb *rxdb,
706 	size_t count)
707 {
708 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
709 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
710 	    BUS_DMASYNC_POSTWRITE);
711 }
712 
713 static void
714 bcmeth_txq_desc_presync(
715 	struct bcmeth_softc *sc,
716 	struct bcmeth_txqueue *txq,
717 	struct gmac_txdb *txdb,
718 	size_t count)
719 {
720 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
721 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
722 	    BUS_DMASYNC_PREWRITE);
723 }
724 
725 static void
726 bcmeth_txq_desc_postsync(
727 	struct bcmeth_softc *sc,
728 	struct bcmeth_txqueue *txq,
729 	struct gmac_txdb *txdb,
730 	size_t count)
731 {
732 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
733 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
734 	    BUS_DMASYNC_POSTWRITE);
735 }
736 
737 static bus_dmamap_t
738 bcmeth_mapcache_get(
739 	struct bcmeth_softc *sc,
740 	struct bcmeth_mapcache *dmc)
741 {
742 	KASSERT(dmc->dmc_nmaps > 0);
743 	KASSERT(dmc->dmc_maps[dmc->dmc_nmaps-1] != NULL);
744 	return dmc->dmc_maps[--dmc->dmc_nmaps];
745 }
746 
747 static void
748 bcmeth_mapcache_put(
749 	struct bcmeth_softc *sc,
750 	struct bcmeth_mapcache *dmc,
751 	bus_dmamap_t map)
752 {
753 	KASSERT(map != NULL);
754 	KASSERT(dmc->dmc_nmaps < dmc->dmc_maxmaps);
755 	dmc->dmc_maps[dmc->dmc_nmaps++] = map;
756 }
757 
758 static void
759 bcmeth_mapcache_destroy(
760 	struct bcmeth_softc *sc,
761 	struct bcmeth_mapcache *dmc)
762 {
763 	const size_t dmc_size =
764 	    offsetof(struct bcmeth_mapcache, dmc_maps[dmc->dmc_maxmaps]);
765 
766 	for (u_int i = 0; i < dmc->dmc_maxmaps; i++) {
767 		bus_dmamap_destroy(sc->sc_dmat, dmc->dmc_maps[i]);
768 	}
769 	kmem_intr_free(dmc, dmc_size);
770 }
771 
772 static int
773 bcmeth_mapcache_create(
774 	struct bcmeth_softc *sc,
775 	struct bcmeth_mapcache **dmc_p,
776 	size_t maxmaps,
777 	size_t maxmapsize,
778 	size_t maxseg)
779 {
780 	const size_t dmc_size =
781 	    offsetof(struct bcmeth_mapcache, dmc_maps[maxmaps]);
782 	struct bcmeth_mapcache * const dmc =
783 		kmem_intr_zalloc(dmc_size, KM_NOSLEEP);
784 
785 	dmc->dmc_maxmaps = maxmaps;
786 	dmc->dmc_nmaps = maxmaps;
787 	dmc->dmc_maxmapsize = maxmapsize;
788 	dmc->dmc_maxseg = maxseg;
789 
790 	for (u_int i = 0; i < maxmaps; i++) {
791 		int error = bus_dmamap_create(sc->sc_dmat, dmc->dmc_maxmapsize,
792 		     dmc->dmc_maxseg, dmc->dmc_maxmapsize, 0,
793 		     BUS_DMA_WAITOK|BUS_DMA_ALLOCNOW, &dmc->dmc_maps[i]);
794 		if (error) {
795 			aprint_error_dev(sc->sc_dev,
796 			    "failed to creat dma map cache "
797 			    "entry %u of %zu: %d\n",
798 			    i, maxmaps, error);
799 			while (i-- > 0) {
800 				bus_dmamap_destroy(sc->sc_dmat,
801 				    dmc->dmc_maps[i]);
802 			}
803 			kmem_intr_free(dmc, dmc_size);
804 			return error;
805 		}
806 		KASSERT(dmc->dmc_maps[i] != NULL);
807 	}
808 
809 	*dmc_p = dmc;
810 
811 	return 0;
812 }
813 
814 #if 0
815 static void
816 bcmeth_dmamem_free(
817 	bus_dma_tag_t dmat,
818 	size_t map_size,
819 	bus_dma_segment_t *seg,
820 	bus_dmamap_t map,
821 	void *kvap)
822 {
823 	bus_dmamap_destroy(dmat, map);
824 	bus_dmamem_unmap(dmat, kvap, map_size);
825 	bus_dmamem_free(dmat, seg, 1);
826 }
827 #endif
828 
829 static int
830 bcmeth_dmamem_alloc(
831 	bus_dma_tag_t dmat,
832 	size_t map_size,
833 	bus_dma_segment_t *seg,
834 	bus_dmamap_t *map,
835 	void **kvap)
836 {
837 	int error;
838 	int nseg;
839 
840 	*kvap = NULL;
841 	*map = NULL;
842 
843 	error = bus_dmamem_alloc(dmat, map_size, 2*PAGE_SIZE, 0,
844 	   seg, 1, &nseg, 0);
845 	if (error)
846 		return error;
847 
848 	KASSERT(nseg == 1);
849 
850 	error = bus_dmamem_map(dmat, seg, nseg, map_size, (void **)kvap, 0);
851 	if (error == 0) {
852 		error = bus_dmamap_create(dmat, map_size, 1, map_size, 0, 0,
853 		    map);
854 		if (error == 0) {
855 			error = bus_dmamap_load(dmat, *map, *kvap, map_size,
856 			    NULL, 0);
857 			if (error == 0)
858 				return 0;
859 			bus_dmamap_destroy(dmat, *map);
860 			*map = NULL;
861 		}
862 		bus_dmamem_unmap(dmat, *kvap, map_size);
863 		*kvap = NULL;
864 	}
865 	bus_dmamem_free(dmat, seg, nseg);
866 	return 0;
867 }
868 
869 static struct mbuf *
870 bcmeth_rx_buf_alloc(
871 	struct bcmeth_softc *sc)
872 {
873 	struct mbuf *m = m_gethdr(M_DONTWAIT, MT_DATA);
874 	if (m == NULL) {
875 		printf("%s:%d: %s\n", __func__, __LINE__, "m_gethdr");
876 		return NULL;
877 	}
878 	MCLGET(m, M_DONTWAIT);
879 	if ((m->m_flags & M_EXT) == 0) {
880 		printf("%s:%d: %s\n", __func__, __LINE__, "MCLGET");
881 		m_freem(m);
882 		return NULL;
883 	}
884 	m->m_len = m->m_pkthdr.len = m->m_ext.ext_size;
885 
886 	bus_dmamap_t map = bcmeth_mapcache_get(sc, sc->sc_rx_mapcache);
887 	if (map == NULL) {
888 		printf("%s:%d: %s\n", __func__, __LINE__, "map get");
889 		m_freem(m);
890 		return NULL;
891 	}
892 	M_SETCTX(m, map);
893 	m->m_len = m->m_pkthdr.len = MCLBYTES;
894 	int error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
895 	    BUS_DMA_READ|BUS_DMA_NOWAIT);
896 	if (error) {
897 		aprint_error_dev(sc->sc_dev, "fail to load rx dmamap: %d\n",
898 		    error);
899 		M_SETCTX(m, NULL);
900 		m_freem(m);
901 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
902 		return NULL;
903 	}
904 	KASSERT(map->dm_mapsize == MCLBYTES);
905 #ifdef BCMETH_RCVMAGIC
906 	*mtod(m, uint32_t *) = htole32(BCMETH_RCVMAGIC);
907 	bus_dmamap_sync(sc->sc_dmat, map, 0, sizeof(uint32_t),
908 	    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
909 	bus_dmamap_sync(sc->sc_dmat, map, sizeof(uint32_t),
910 	    map->dm_mapsize - sizeof(uint32_t), BUS_DMASYNC_PREREAD);
911 #else
912 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
913 	    BUS_DMASYNC_PREREAD);
914 #endif
915 
916 	return m;
917 }
918 
919 static void
920 bcmeth_rx_map_unload(
921 	struct bcmeth_softc *sc,
922 	struct mbuf *m)
923 {
924 	KASSERT(m);
925 	for (; m != NULL; m = m->m_next) {
926 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
927 		KASSERT(map);
928 		KASSERT(map->dm_mapsize == MCLBYTES);
929 		bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_len,
930 		    BUS_DMASYNC_POSTREAD);
931 		bus_dmamap_unload(sc->sc_dmat, map);
932 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
933 		M_SETCTX(m, NULL);
934 	}
935 }
936 
937 static bool
938 bcmeth_rxq_produce(
939 	struct bcmeth_softc *sc,
940 	struct bcmeth_rxqueue *rxq)
941 {
942 	struct gmac_rxdb *producer = rxq->rxq_producer;
943 	bool produced = false;
944 
945 	while (rxq->rxq_inuse < rxq->rxq_threshold) {
946 		struct mbuf *m;
947 		IF_DEQUEUE(&sc->sc_rx_bufcache, m);
948 		if (m == NULL) {
949 			m = bcmeth_rx_buf_alloc(sc);
950 			if (m == NULL) {
951 				printf("%s: bcmeth_rx_buf_alloc failed\n", __func__);
952 				break;
953 			}
954 		}
955 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
956 		KASSERT(map);
957 
958 		producer->rxdb_buflen = htole32(MCLBYTES);
959 		producer->rxdb_addrlo = htole32(map->dm_segs[0].ds_addr);
960 		producer->rxdb_flags &= htole32(RXDB_FLAG_ET);
961 		*rxq->rxq_mtail = m;
962 		rxq->rxq_mtail = &m->m_next;
963 		m->m_len = MCLBYTES;
964 		m->m_next = NULL;
965 		rxq->rxq_inuse++;
966 		if (++producer == rxq->rxq_last) {
967 			membar_producer();
968 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
969 			    rxq->rxq_last - rxq->rxq_producer);
970 			producer = rxq->rxq_producer = rxq->rxq_first;
971 		}
972 		produced = true;
973 	}
974 	if (produced) {
975 		membar_producer();
976 		if (producer != rxq->rxq_producer) {
977 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
978 			    producer - rxq->rxq_producer);
979 			rxq->rxq_producer = producer;
980 		}
981 		bcmeth_write_4(sc, rxq->rxq_reg_rcvptr,
982 		    rxq->rxq_descmap->dm_segs[0].ds_addr
983 		    + ((uintptr_t)producer & RCVPTR));
984 	}
985 	return true;
986 }
987 
988 static void
989 bcmeth_rx_input(
990 	struct bcmeth_softc *sc,
991 	struct mbuf *m,
992 	uint32_t rxdb_flags)
993 {
994 	struct ifnet * const ifp = &sc->sc_if;
995 
996 	bcmeth_rx_map_unload(sc, m);
997 
998 	m_adj(m, sc->sc_rcvoffset);
999 
1000 	/*
1001 	 * If we are in promiscuous mode and this isn't a multicast, check the
1002 	 * destination address to make sure it matches our own.  If it doesn't,
1003 	 * mark the packet as being received promiscuously.
1004 	 */
1005 	if ((sc->sc_cmdcfg & PROMISC_EN)
1006 	    && (m->m_data[0] & 1) == 0
1007 	    && (*(uint16_t *)&m->m_data[0] != sc->sc_macaddr[0]
1008 		|| *(uint32_t *)&m->m_data[2] != sc->sc_macaddr[1])) {
1009 		m->m_flags |= M_PROMISC;
1010 	}
1011 	m->m_pkthdr.rcvif = ifp;
1012 
1013 	ifp->if_ipackets++;
1014 	ifp->if_ibytes += m->m_pkthdr.len;
1015 
1016 	/*
1017 	 * Let's give it to the network subsystm to deal with.
1018 	 */
1019 #ifdef BCMETH_MPSAFE
1020 	mutex_exit(sc->sc_lock);
1021 	(*ifp->if_input)(ifp, m);
1022 	mutex_enter(sc->sc_lock);
1023 #else
1024 	int s = splnet();
1025 	bpf_mtap(ifp, m);
1026 	(*ifp->if_input)(ifp, m);
1027 	splx(s);
1028 #endif
1029 }
1030 
1031 static bool
1032 bcmeth_rxq_consume(
1033 	struct bcmeth_softc *sc,
1034 	struct bcmeth_rxqueue *rxq,
1035 	size_t atmost)
1036 {
1037 	struct ifnet * const ifp = &sc->sc_if;
1038 	struct gmac_rxdb *consumer = rxq->rxq_consumer;
1039 	size_t rxconsumed = 0;
1040 	bool didconsume = false;
1041 
1042 	while (atmost-- > 0) {
1043 		if (consumer == rxq->rxq_producer) {
1044 			KASSERT(rxq->rxq_inuse == 0);
1045 			break;
1046 		}
1047 
1048 		uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1049 		uint32_t currdscr = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1050 		if (consumer == rxq->rxq_first + currdscr) {
1051 			break;
1052 		}
1053 		bcmeth_rxq_desc_postsync(sc, rxq, consumer, 1);
1054 
1055 		/*
1056 		 * We own this packet again.  Copy the rxsts word from it.
1057 		 */
1058 		rxconsumed++;
1059 		didconsume = true;
1060 		uint32_t rxsts;
1061 		KASSERT(rxq->rxq_mhead != NULL);
1062 		bus_dmamap_t map = M_GETCTX(rxq->rxq_mhead, bus_dmamap_t);
1063 		bus_dmamap_sync(sc->sc_dmat, map, 0, arm_dcache_align,
1064 		    BUS_DMASYNC_POSTREAD);
1065 		memcpy(&rxsts, rxq->rxq_mhead->m_data, 4);
1066 		rxsts = le32toh(rxsts);
1067 #if 0
1068 		KASSERTMSG(rxsts != BCMETH_RCVMAGIC, "currdscr=%u consumer=%zd",
1069 		    currdscr, consumer - rxq->rxq_first);
1070 #endif
1071 
1072 		/*
1073 		 * Get the count of descriptors.  Fetch the correct number
1074 		 * of mbufs.
1075 		 */
1076 #ifdef BCMETH_RCVMAGIC
1077 		size_t desc_count = rxsts != BCMETH_RCVMAGIC ? __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1 : 1;
1078 #else
1079 		size_t desc_count = __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1;
1080 #endif
1081 		struct mbuf *m = rxq->rxq_mhead;
1082 		struct mbuf *m_last = m;
1083 		for (size_t i = 1; i < desc_count; i++) {
1084 			if (++consumer == rxq->rxq_last) {
1085 				consumer = rxq->rxq_first;
1086 			}
1087 			KASSERTMSG(consumer != rxq->rxq_first + currdscr,
1088 			    "i=%zu rxsts=%#x desc_count=%zu currdscr=%u consumer=%zd",
1089 			    i, rxsts, desc_count, currdscr,
1090 			    consumer - rxq->rxq_first);
1091 			m_last = m_last->m_next;
1092 		}
1093 
1094 		/*
1095 		 * Now remove it/them from the list of enqueued mbufs.
1096 		 */
1097 		if ((rxq->rxq_mhead = m_last->m_next) == NULL)
1098 			rxq->rxq_mtail = &rxq->rxq_mhead;
1099 		m_last->m_next = NULL;
1100 
1101 #ifdef BCMETH_RCVMAGIC
1102 		if (rxsts == BCMETH_RCVMAGIC) {
1103 			ifp->if_ierrors++;
1104 			if ((m->m_ext.ext_paddr >> 28) == 8) {
1105 				BCMETH_EVCNT_INCR(sc->sc_ev_rx_badmagic_lo);
1106 			} else {
1107 				BCMETH_EVCNT_INCR( sc->sc_ev_rx_badmagic_hi);
1108 			}
1109 			IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1110 		} else
1111 #endif /* BCMETH_RCVMAGIC */
1112 		if (rxsts & (RXSTS_CRC_ERROR|RXSTS_OVERSIZED|RXSTS_PKT_OVERFLOW)) {
1113 			aprint_error_dev(sc->sc_dev, "[%zu]: count=%zu rxsts=%#x\n",
1114 			    consumer - rxq->rxq_first, desc_count, rxsts);
1115 			/*
1116 			 * We encountered an error, take the mbufs and add them
1117 			 * to the rx bufcache so we can quickly reuse them.
1118 			 */
1119 			ifp->if_ierrors++;
1120 			do {
1121 				struct mbuf *m0 = m->m_next;
1122 				m->m_next = NULL;
1123 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1124 				m = m0;
1125 			} while (m);
1126 		} else {
1127 			uint32_t framelen = __SHIFTOUT(rxsts, RXSTS_FRAMELEN);
1128 			framelen += sc->sc_rcvoffset;
1129 			m->m_pkthdr.len = framelen;
1130 			if (desc_count == 1) {
1131 				KASSERT(framelen <= MCLBYTES);
1132 				m->m_len = framelen;
1133 			} else {
1134 				m_last->m_len = framelen & (MCLBYTES - 1);
1135 			}
1136 
1137 #ifdef BCMETH_MPSAFE
1138 			/*
1139 			 * Wrap at the last entry!
1140 			 */
1141 			if (++consumer == rxq->rxq_last) {
1142 				KASSERT(consumer[-1].rxdb_flags & htole32(RXDB_FLAG_ET));
1143 				rxq->rxq_consumer = rxq->rxq_first;
1144 			} else {
1145 				rxq->rxq_consumer = consumer;
1146 			}
1147 			rxq->rxq_inuse -= rxconsumed;
1148 #endif /* BCMETH_MPSAFE */
1149 
1150 			/*
1151 			 * Receive the packet (which releases our lock)
1152 			 */
1153 			bcmeth_rx_input(sc, m, rxsts);
1154 
1155 #ifdef BCMETH_MPSAFE
1156 			/*
1157 			 * Since we had to give up our lock, we need to
1158 			 * refresh these.
1159 			 */
1160 			consumer = rxq->rxq_consumer;
1161 			rxconsumed = 0;
1162 			continue;
1163 #endif /* BCMETH_MPSAFE */
1164 		}
1165 
1166 		/*
1167 		 * Wrap at the last entry!
1168 		 */
1169 		if (++consumer == rxq->rxq_last) {
1170 			KASSERT(consumer[-1].rxdb_flags & htole32(RXDB_FLAG_ET));
1171 			consumer = rxq->rxq_first;
1172 		}
1173 	}
1174 
1175 	/*
1176 	 * Update queue info.
1177 	 */
1178 	rxq->rxq_consumer = consumer;
1179 	rxq->rxq_inuse -= rxconsumed;
1180 
1181 	/*
1182 	 * Did we consume anything?
1183 	 */
1184 	return didconsume;
1185 }
1186 
1187 static void
1188 bcmeth_rxq_purge(
1189 	struct bcmeth_softc *sc,
1190 	struct bcmeth_rxqueue *rxq,
1191 	bool discard)
1192 {
1193 	struct mbuf *m;
1194 
1195 	if ((m = rxq->rxq_mhead) != NULL) {
1196 		if (discard) {
1197 			bcmeth_rx_map_unload(sc, m);
1198 			m_freem(m);
1199 		} else {
1200 			while (m != NULL) {
1201 				struct mbuf *m0 = m->m_next;
1202 				m->m_next = NULL;
1203 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1204 				m = m0;
1205 			}
1206 		}
1207 
1208 	}
1209 
1210 	rxq->rxq_mhead = NULL;
1211 	rxq->rxq_mtail = &rxq->rxq_mhead;
1212 	rxq->rxq_inuse = 0;
1213 }
1214 
1215 static void
1216 bcmeth_rxq_reset(
1217 	struct bcmeth_softc *sc,
1218 	struct bcmeth_rxqueue *rxq)
1219 {
1220 	/*
1221 	 * sync all the descriptors
1222 	 */
1223 	bcmeth_rxq_desc_postsync(sc, rxq, rxq->rxq_first,
1224 	    rxq->rxq_last - rxq->rxq_first);
1225 
1226 	/*
1227 	 * Make sure we own all descriptors in the ring.
1228 	 */
1229 	struct gmac_rxdb *rxdb;
1230 	for (rxdb = rxq->rxq_first; rxdb < rxq->rxq_last - 1; rxdb++) {
1231 		rxdb->rxdb_flags = htole32(RXDB_FLAG_IC);
1232 	}
1233 
1234 	/*
1235 	 * Last descriptor has the wrap flag.
1236 	 */
1237 	rxdb->rxdb_flags = htole32(RXDB_FLAG_ET|RXDB_FLAG_IC);
1238 
1239 	/*
1240 	 * Reset the producer consumer indexes.
1241 	 */
1242 	rxq->rxq_consumer = rxq->rxq_first;
1243 	rxq->rxq_producer = rxq->rxq_first;
1244 	rxq->rxq_inuse = 0;
1245 	if (rxq->rxq_threshold < BCMETH_MINRXMBUFS)
1246 		rxq->rxq_threshold = BCMETH_MINRXMBUFS;
1247 
1248 	sc->sc_intmask |= RCVINT|RCVFIFOOF|RCVDESCUF;
1249 
1250 	/*
1251 	 * Restart the receiver at the first descriptor
1252 	 */
1253 	bcmeth_write_4(sc, rxq->rxq_reg_rcvaddrlo,
1254 	    rxq->rxq_descmap->dm_segs[0].ds_addr);
1255 }
1256 
1257 static int
1258 bcmeth_rxq_attach(
1259 	struct bcmeth_softc *sc,
1260 	struct bcmeth_rxqueue *rxq,
1261 	u_int qno)
1262 {
1263 	size_t desc_count = BCMETH_RINGSIZE / sizeof(rxq->rxq_first[0]);
1264 	int error;
1265 	void *descs;
1266 
1267 	KASSERT(desc_count == 256 || desc_count == 512);
1268 
1269 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1270 	   &rxq->rxq_descmap_seg, &rxq->rxq_descmap, &descs);
1271 	if (error)
1272 		return error;
1273 
1274 	memset(descs, 0, BCMETH_RINGSIZE);
1275 	rxq->rxq_first = descs;
1276 	rxq->rxq_last = rxq->rxq_first + desc_count;
1277 	rxq->rxq_consumer = descs;
1278 	rxq->rxq_producer = descs;
1279 
1280 	bcmeth_rxq_purge(sc, rxq, true);
1281 	bcmeth_rxq_reset(sc, rxq);
1282 
1283 	rxq->rxq_reg_rcvaddrlo = GMAC_RCVADDR_LOW;
1284 	rxq->rxq_reg_rcvctl = GMAC_RCVCONTROL;
1285 	rxq->rxq_reg_rcvptr = GMAC_RCVPTR;
1286 	rxq->rxq_reg_rcvsts0 = GMAC_RCVSTATUS0;
1287 	rxq->rxq_reg_rcvsts1 = GMAC_RCVSTATUS1;
1288 
1289 	return 0;
1290 }
1291 
1292 static bool
1293 bcmeth_txq_active_p(
1294 	struct bcmeth_softc * const sc,
1295 	struct bcmeth_txqueue *txq)
1296 {
1297 	return !IF_IS_EMPTY(&txq->txq_mbufs);
1298 }
1299 
1300 static bool
1301 bcmeth_txq_fillable_p(
1302 	struct bcmeth_softc * const sc,
1303 	struct bcmeth_txqueue *txq)
1304 {
1305 	return txq->txq_free >= txq->txq_threshold;
1306 }
1307 
1308 static int
1309 bcmeth_txq_attach(
1310 	struct bcmeth_softc *sc,
1311 	struct bcmeth_txqueue *txq,
1312 	u_int qno)
1313 {
1314 	size_t desc_count = BCMETH_RINGSIZE / sizeof(txq->txq_first[0]);
1315 	int error;
1316 	void *descs;
1317 
1318 	KASSERT(desc_count == 256 || desc_count == 512);
1319 
1320 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1321 	   &txq->txq_descmap_seg, &txq->txq_descmap, &descs);
1322 	if (error)
1323 		return error;
1324 
1325 	memset(descs, 0, BCMETH_RINGSIZE);
1326 	txq->txq_first = descs;
1327 	txq->txq_last = txq->txq_first + desc_count;
1328 	txq->txq_consumer = descs;
1329 	txq->txq_producer = descs;
1330 
1331 	IFQ_SET_MAXLEN(&txq->txq_mbufs, BCMETH_MAXTXMBUFS);
1332 
1333 	txq->txq_reg_xmtaddrlo = GMAC_XMTADDR_LOW;
1334 	txq->txq_reg_xmtctl = GMAC_XMTCONTROL;
1335 	txq->txq_reg_xmtptr = GMAC_XMTPTR;
1336 	txq->txq_reg_xmtsts0 = GMAC_XMTSTATUS0;
1337 	txq->txq_reg_xmtsts1 = GMAC_XMTSTATUS1;
1338 
1339 	bcmeth_txq_reset(sc, txq);
1340 
1341 	return 0;
1342 }
1343 
1344 static int
1345 bcmeth_txq_map_load(
1346 	struct bcmeth_softc *sc,
1347 	struct bcmeth_txqueue *txq,
1348 	struct mbuf *m)
1349 {
1350 	bus_dmamap_t map;
1351 	int error;
1352 
1353 	map = M_GETCTX(m, bus_dmamap_t);
1354 	if (map != NULL)
1355 		return 0;
1356 
1357 	map = bcmeth_mapcache_get(sc, sc->sc_tx_mapcache);
1358 	if (map == NULL)
1359 		return ENOMEM;
1360 
1361 	error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
1362 	    BUS_DMA_WRITE | BUS_DMA_NOWAIT);
1363 	if (error)
1364 		return error;
1365 
1366 	bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_pkthdr.len,
1367 	    BUS_DMASYNC_PREWRITE);
1368 	M_SETCTX(m, map);
1369 	return 0;
1370 }
1371 
1372 static void
1373 bcmeth_txq_map_unload(
1374 	struct bcmeth_softc *sc,
1375 	struct bcmeth_txqueue *txq,
1376 	struct mbuf *m)
1377 {
1378 	KASSERT(m);
1379 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1380 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1381 	    BUS_DMASYNC_POSTWRITE);
1382 	bus_dmamap_unload(sc->sc_dmat, map);
1383 	bcmeth_mapcache_put(sc, sc->sc_tx_mapcache, map);
1384 }
1385 
1386 static bool
1387 bcmeth_txq_produce(
1388 	struct bcmeth_softc *sc,
1389 	struct bcmeth_txqueue *txq,
1390 	struct mbuf *m)
1391 {
1392 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1393 
1394 	if (map->dm_nsegs > txq->txq_free)
1395 		return false;
1396 
1397 	/*
1398 	 * TCP Offload flag must be set in the first descriptor.
1399 	 */
1400 	struct gmac_txdb *producer = txq->txq_producer;
1401 	uint32_t first_flags = TXDB_FLAG_SF;
1402 	uint32_t last_flags = TXDB_FLAG_EF;
1403 
1404 	/*
1405 	 * If we've produced enough descriptors without consuming any
1406 	 * we need to ask for an interrupt to reclaim some.
1407 	 */
1408 	txq->txq_lastintr += map->dm_nsegs;
1409 	if (txq->txq_lastintr >= txq->txq_threshold
1410 	    || txq->txq_mbufs.ifq_len + 1 == txq->txq_mbufs.ifq_maxlen) {
1411 		txq->txq_lastintr = 0;
1412 		last_flags |= TXDB_FLAG_IC;
1413 	}
1414 
1415 	KASSERT(producer != txq->txq_last);
1416 
1417 	struct gmac_txdb *start = producer;
1418 	size_t count = map->dm_nsegs;
1419 	producer->txdb_flags |= htole32(first_flags);
1420 	producer->txdb_addrlo = htole32(map->dm_segs[0].ds_addr);
1421 	producer->txdb_buflen = htole32(map->dm_segs[0].ds_len);
1422 	for (u_int i = 1; i < map->dm_nsegs; i++) {
1423 #if 0
1424 		printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1425 		    le32toh(producer->txdb_flags),
1426 		    le32toh(producer->txdb_buflen),
1427 		    le32toh(producer->txdb_addrlo),
1428 		    le32toh(producer->txdb_addrhi));
1429 #endif
1430 		if (__predict_false(++producer == txq->txq_last)) {
1431 			bcmeth_txq_desc_presync(sc, txq, start,
1432 			    txq->txq_last - start);
1433 			count -= txq->txq_last - start;
1434 			producer = txq->txq_first;
1435 			start = txq->txq_first;
1436 		}
1437 		producer->txdb_addrlo = htole32(map->dm_segs[i].ds_addr);
1438 		producer->txdb_buflen = htole32(map->dm_segs[i].ds_len);
1439 	}
1440 	producer->txdb_flags |= htole32(last_flags);
1441 #if 0
1442 	printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1443 	    le32toh(producer->txdb_flags), le32toh(producer->txdb_buflen),
1444 	    le32toh(producer->txdb_addrlo), le32toh(producer->txdb_addrhi));
1445 #endif
1446 	if (count)
1447 		bcmeth_txq_desc_presync(sc, txq, start, count);
1448 
1449 	/*
1450 	 * Reduce free count by the number of segments we consumed.
1451 	 */
1452 	txq->txq_free -= map->dm_nsegs;
1453 	KASSERT(map->dm_nsegs == 1 || txq->txq_producer != producer);
1454 	KASSERT(map->dm_nsegs == 1 || (txq->txq_producer->txdb_flags & htole32(TXDB_FLAG_EF)) == 0);
1455 	KASSERT(producer->txdb_flags & htole32(TXDB_FLAG_EF));
1456 
1457 #if 0
1458 	printf("%s: mbuf %p: produced a %u byte packet in %u segments (%zd..%zd)\n",
1459 	    __func__, m, m->m_pkthdr.len, map->dm_nsegs,
1460 	    txq->txq_producer - txq->txq_first, producer - txq->txq_first);
1461 #endif
1462 
1463 	if (producer + 1 == txq->txq_last)
1464 		txq->txq_producer = txq->txq_first;
1465 	else
1466 		txq->txq_producer = producer + 1;
1467 	IF_ENQUEUE(&txq->txq_mbufs, m);
1468 
1469 	/*
1470 	 * Let the transmitter know there's more to do
1471 	 */
1472 	bcmeth_write_4(sc, txq->txq_reg_xmtptr,
1473 	    txq->txq_descmap->dm_segs[0].ds_addr
1474 	    + ((uintptr_t)txq->txq_producer & XMT_LASTDSCR));
1475 
1476 	return true;
1477 }
1478 
1479 static struct mbuf *
1480 bcmeth_copy_packet(struct mbuf *m)
1481 {
1482 	struct mbuf *mext = NULL;
1483 	size_t misalignment = 0;
1484 	size_t hlen = 0;
1485 
1486 	for (mext = m; mext != NULL; mext = mext->m_next) {
1487 		if (mext->m_flags & M_EXT) {
1488 			misalignment = mtod(mext, vaddr_t) & arm_dcache_align;
1489 			break;
1490 		}
1491 		hlen += m->m_len;
1492 	}
1493 
1494 	struct mbuf *n = m->m_next;
1495 	if (m != mext && hlen + misalignment <= MHLEN && false) {
1496 		KASSERT(m->m_pktdat <= m->m_data && m->m_data <= &m->m_pktdat[MHLEN - m->m_len]);
1497 		size_t oldoff = m->m_data - m->m_pktdat;
1498 		size_t off;
1499 		if (mext == NULL) {
1500 			off = (oldoff + hlen > MHLEN) ? 0 : oldoff;
1501 		} else {
1502 			off = MHLEN - (hlen + misalignment);
1503 		}
1504 		KASSERT(off + hlen + misalignment <= MHLEN);
1505 		if (((oldoff ^ off) & arm_dcache_align) != 0 || off < oldoff) {
1506 			memmove(&m->m_pktdat[off], m->m_data, m->m_len);
1507 			m->m_data = &m->m_pktdat[off];
1508 		}
1509 		m_copydata(n, 0, hlen - m->m_len, &m->m_data[m->m_len]);
1510 		m->m_len = hlen;
1511 		m->m_next = mext;
1512 		while (n != mext) {
1513 			n = m_free(n);
1514 		}
1515 		return m;
1516 	}
1517 
1518 	struct mbuf *m0 = m_gethdr(M_DONTWAIT, m->m_type);
1519 	if (m0 == NULL) {
1520 		return NULL;
1521 	}
1522 	M_COPY_PKTHDR(m0, m);
1523 	MCLAIM(m0, m->m_owner);
1524 	if (m0->m_pkthdr.len > MHLEN) {
1525 		MCLGET(m0, M_DONTWAIT);
1526 		if ((m0->m_flags & M_EXT) == 0) {
1527 			m_freem(m0);
1528 			return NULL;
1529 		}
1530 	}
1531 	m0->m_len = m->m_pkthdr.len;
1532 	m_copydata(m, 0, m0->m_len, mtod(m0, void *));
1533 	m_freem(m);
1534 	return m0;
1535 }
1536 
1537 static bool
1538 bcmeth_txq_enqueue(
1539 	struct bcmeth_softc *sc,
1540 	struct bcmeth_txqueue *txq)
1541 {
1542 	for (;;) {
1543 		if (IF_QFULL(&txq->txq_mbufs))
1544 			return false;
1545 		struct mbuf *m = txq->txq_next;
1546 		if (m == NULL) {
1547 			int s = splnet();
1548 			IF_DEQUEUE(&sc->sc_if.if_snd, m);
1549 			splx(s);
1550 			if (m == NULL)
1551 				return true;
1552 			M_SETCTX(m, NULL);
1553 		} else {
1554 			txq->txq_next = NULL;
1555 		}
1556 		/*
1557 		 * If LINK2 is set and this packet uses multiple mbufs,
1558 		 * consolidate it into a single mbuf.
1559 		 */
1560 		if (m->m_next != NULL && (sc->sc_if.if_flags & IFF_LINK2)) {
1561 			struct mbuf *m0 = bcmeth_copy_packet(m);
1562 			if (m0 == NULL) {
1563 				txq->txq_next = m;
1564 				return true;
1565 			}
1566 			m = m0;
1567 		}
1568 		int error = bcmeth_txq_map_load(sc, txq, m);
1569 		if (error) {
1570 			aprint_error_dev(sc->sc_dev,
1571 			    "discarded packet due to "
1572 			    "dmamap load failure: %d\n", error);
1573 			m_freem(m);
1574 			continue;
1575 		}
1576 		KASSERT(txq->txq_next == NULL);
1577 		if (!bcmeth_txq_produce(sc, txq, m)) {
1578 			txq->txq_next = m;
1579 			return false;
1580 		}
1581 		KASSERT(txq->txq_next == NULL);
1582 	}
1583 }
1584 
1585 static bool
1586 bcmeth_txq_consume(
1587 	struct bcmeth_softc *sc,
1588 	struct bcmeth_txqueue *txq)
1589 {
1590 	struct ifnet * const ifp = &sc->sc_if;
1591 	struct gmac_txdb *consumer = txq->txq_consumer;
1592 	size_t txfree = 0;
1593 
1594 #if 0
1595 	printf("%s: entry: free=%zu\n", __func__, txq->txq_free);
1596 #endif
1597 
1598 	for (;;) {
1599 		if (consumer == txq->txq_producer) {
1600 			txq->txq_consumer = consumer;
1601 			txq->txq_free += txfree;
1602 			txq->txq_lastintr -= min(txq->txq_lastintr, txfree);
1603 #if 0
1604 			printf("%s: empty: freed %zu descriptors going from %zu to %zu\n",
1605 			    __func__, txfree, txq->txq_free - txfree, txq->txq_free);
1606 #endif
1607 			KASSERT(txq->txq_lastintr == 0);
1608 			KASSERT(txq->txq_free == txq->txq_last - txq->txq_first - 1);
1609 			return true;
1610 		}
1611 		bcmeth_txq_desc_postsync(sc, txq, consumer, 1);
1612 		uint32_t s0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
1613 		if (consumer == txq->txq_first + __SHIFTOUT(s0, XMT_CURRDSCR)) {
1614 			txq->txq_consumer = consumer;
1615 			txq->txq_free += txfree;
1616 			txq->txq_lastintr -= min(txq->txq_lastintr, txfree);
1617 #if 0
1618 			printf("%s: freed %zu descriptors\n",
1619 			    __func__, txfree);
1620 #endif
1621 			return bcmeth_txq_fillable_p(sc, txq);
1622 		}
1623 
1624 		/*
1625 		 * If this is the last descriptor in the chain, get the
1626 		 * mbuf, free its dmamap, and free the mbuf chain itself.
1627 		 */
1628 		const uint32_t txdb_flags = le32toh(consumer->txdb_flags);
1629 		if (txdb_flags & TXDB_FLAG_EF) {
1630 			struct mbuf *m;
1631 
1632 			IF_DEQUEUE(&txq->txq_mbufs, m);
1633 			KASSERT(m);
1634 			bcmeth_txq_map_unload(sc, txq, m);
1635 #if 0
1636 			printf("%s: mbuf %p: consumed a %u byte packet\n",
1637 			    __func__, m, m->m_pkthdr.len);
1638 #endif
1639 			bpf_mtap(ifp, m);
1640 			ifp->if_opackets++;
1641 			ifp->if_obytes += m->m_pkthdr.len;
1642 			if (m->m_flags & M_MCAST)
1643 				ifp->if_omcasts++;
1644 			m_freem(m);
1645 		}
1646 
1647 		/*
1648 		 * We own this packet again.  Clear all flags except wrap.
1649 		 */
1650 		txfree++;
1651 
1652 		/*
1653 		 * Wrap at the last entry!
1654 		 */
1655 		if (txdb_flags & TXDB_FLAG_ET) {
1656 			consumer->txdb_flags = htole32(TXDB_FLAG_ET);
1657 			KASSERT(consumer + 1 == txq->txq_last);
1658 			consumer = txq->txq_first;
1659 		} else {
1660 			consumer->txdb_flags = 0;
1661 			consumer++;
1662 			KASSERT(consumer < txq->txq_last);
1663 		}
1664 	}
1665 }
1666 
1667 static void
1668 bcmeth_txq_purge(
1669 	struct bcmeth_softc *sc,
1670 	struct bcmeth_txqueue *txq)
1671 {
1672 	struct mbuf *m;
1673 	KASSERT((bcmeth_read_4(sc, UNIMAC_COMMAND_CONFIG) & TX_ENA) == 0);
1674 
1675 	for (;;) {
1676 		IF_DEQUEUE(&txq->txq_mbufs, m);
1677 		if (m == NULL)
1678 			break;
1679 		bcmeth_txq_map_unload(sc, txq, m);
1680 		m_freem(m);
1681 	}
1682 	if ((m = txq->txq_next) != NULL) {
1683 		txq->txq_next = NULL;
1684 		bcmeth_txq_map_unload(sc, txq, m);
1685 		m_freem(m);
1686 	}
1687 }
1688 
1689 static void
1690 bcmeth_txq_reset(
1691 	struct bcmeth_softc *sc,
1692 	struct bcmeth_txqueue *txq)
1693 {
1694 	/*
1695 	 * sync all the descriptors
1696 	 */
1697 	bcmeth_txq_desc_postsync(sc, txq, txq->txq_first,
1698 	    txq->txq_last - txq->txq_first);
1699 
1700 	/*
1701 	 * Make sure we own all descriptors in the ring.
1702 	 */
1703 	struct gmac_txdb *txdb;
1704 	for (txdb = txq->txq_first; txdb < txq->txq_last - 1; txdb++) {
1705 		txdb->txdb_flags = 0;
1706 	}
1707 
1708 	/*
1709 	 * Last descriptor has the wrap flag.
1710 	 */
1711 	txdb->txdb_flags = htole32(TXDB_FLAG_ET);
1712 
1713 	/*
1714 	 * Reset the producer consumer indexes.
1715 	 */
1716 	txq->txq_consumer = txq->txq_first;
1717 	txq->txq_producer = txq->txq_first;
1718 	txq->txq_free = txq->txq_last - txq->txq_first - 1;
1719 	txq->txq_threshold = txq->txq_free / 2;
1720 	txq->txq_lastintr = 0;
1721 
1722 	/*
1723 	 * What do we want to get interrupted on?
1724 	 */
1725 	sc->sc_intmask |= XMTINT_0 | XMTUF;
1726 
1727 	/*
1728 	 * Restart the transmiter at the first descriptor
1729 	 */
1730 	bcmeth_write_4(sc, txq->txq_reg_xmtaddrlo,
1731 	    txq->txq_descmap->dm_segs->ds_addr);
1732 }
1733 
1734 static void
1735 bcmeth_ifstart(struct ifnet *ifp)
1736 {
1737 	struct bcmeth_softc * const sc = ifp->if_softc;
1738 
1739 	if (__predict_false((ifp->if_flags & IFF_RUNNING) == 0)) {
1740 		return;
1741 	}
1742 
1743 #ifdef BCMETH_MPSAFETX
1744 	if (cpu_intr_p()) {
1745 #endif
1746 		atomic_or_uint(&sc->sc_soft_flags, SOFT_TXINTR);
1747 		softint_schedule(sc->sc_soft_ih);
1748 #ifdef BCMETH_MPSAFETX
1749 	} else {
1750 		/*
1751 		 * Either we are in a softintr thread already or some other
1752 		 * thread so just borrow it to do the send and save ourselves
1753 		 * the overhead of a fast soft int.
1754 		 */
1755 		bcmeth_soft_txintr(sc);
1756 	}
1757 #endif
1758 }
1759 
1760 int
1761 bcmeth_intr(void *arg)
1762 {
1763 	struct bcmeth_softc * const sc = arg;
1764 	uint32_t soft_flags = 0;
1765 	uint32_t work_flags = 0;
1766 	int rv = 0;
1767 
1768 	mutex_enter(sc->sc_hwlock);
1769 
1770 	uint32_t intmask = sc->sc_intmask;
1771 	BCMETH_EVCNT_INCR(sc->sc_ev_intr);
1772 
1773 	for (;;) {
1774 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
1775 		intstatus &= intmask;
1776 		bcmeth_write_4(sc, GMAC_INTSTATUS, intstatus);	/* write 1 to clear */
1777 		if (intstatus == 0) {
1778 			break;
1779 		}
1780 #if 0
1781 		aprint_normal_dev(sc->sc_dev, "%s: intstatus=%#x intmask=%#x\n",
1782 		    __func__, intstatus, bcmeth_read_4(sc, GMAC_INTMASK));
1783 #endif
1784 		if (intstatus & RCVINT) {
1785 			struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
1786 			intmask &= ~RCVINT;
1787 
1788 			uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1789 			uint32_t descs = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1790 			if (descs < rxq->rxq_consumer - rxq->rxq_first) {
1791 				/*
1792 				 * We wrapped at the end so count how far
1793 				 * we are from the end.
1794 				 */
1795 				descs += rxq->rxq_last - rxq->rxq_consumer;
1796 			} else {
1797 				descs -= rxq->rxq_consumer - rxq->rxq_first;
1798 			}
1799 			/*
1800 			 * If we "timedout" we can't be hogging so use
1801 			 * softints.  If we exceeded then we might hogging
1802 			 * so let the workqueue deal with them.
1803 			 */
1804 			const uint32_t framecount = __SHIFTOUT(sc->sc_rcvlazy, INTRCVLAZY_FRAMECOUNT);
1805 			if (descs < framecount
1806 			    || (curcpu()->ci_curlwp->l_flag & LW_IDLE)) {
1807 				soft_flags |= SOFT_RXINTR;
1808 			} else {
1809 				work_flags |= WORK_RXINTR;
1810 			}
1811 		}
1812 
1813 		if (intstatus & XMTINT_0) {
1814 			intmask &= ~XMTINT_0;
1815 			soft_flags |= SOFT_TXINTR;
1816 		}
1817 
1818 		if (intstatus & RCVDESCUF) {
1819 			intmask &= ~RCVDESCUF;
1820 			work_flags |= WORK_RXUNDERFLOW;
1821 		}
1822 
1823 		intstatus &= intmask;
1824 		if (intstatus) {
1825 			aprint_error_dev(sc->sc_dev,
1826 			    "intr: intstatus=%#x\n", intstatus);
1827 			aprint_error_dev(sc->sc_dev,
1828 			    "rcvbase=%p/%#lx rcvptr=%#x rcvsts=%#x/%#x\n",
1829 			    sc->sc_rxq.rxq_first,
1830 			    sc->sc_rxq.rxq_descmap->dm_segs[0].ds_addr,
1831 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvptr),
1832 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts0),
1833 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts1));
1834 			aprint_error_dev(sc->sc_dev,
1835 			    "xmtbase=%p/%#lx xmtptr=%#x xmtsts=%#x/%#x\n",
1836 			    sc->sc_txq.txq_first,
1837 			    sc->sc_txq.txq_descmap->dm_segs[0].ds_addr,
1838 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtptr),
1839 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts0),
1840 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts1));
1841 			intmask &= ~intstatus;
1842 			work_flags |= WORK_REINIT;
1843 			break;
1844 		}
1845 	}
1846 
1847 	if (intmask != sc->sc_intmask) {
1848 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1849 	}
1850 
1851 	if (work_flags) {
1852 		if (sc->sc_work_flags == 0) {
1853 			workqueue_enqueue(sc->sc_workq, &sc->sc_work, NULL);
1854 		}
1855 		atomic_or_32(&sc->sc_work_flags, work_flags);
1856 		rv = 1;
1857 	}
1858 
1859 	if (soft_flags) {
1860 		if (sc->sc_soft_flags == 0) {
1861 			softint_schedule(sc->sc_soft_ih);
1862 		}
1863 		atomic_or_32(&sc->sc_soft_flags, soft_flags);
1864 		rv = 1;
1865 	}
1866 
1867 	mutex_exit(sc->sc_hwlock);
1868 
1869 	return rv;
1870 }
1871 
1872 #ifdef BCMETH_MPSAFETX
1873 void
1874 bcmeth_soft_txintr(struct bcmeth_softc *sc)
1875 {
1876 	mutex_enter(sc->sc_lock);
1877 	/*
1878 	 * Let's do what we came here for.  Consume transmitted
1879 	 * packets off the the transmit ring.
1880 	 */
1881 	if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1882 	    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1883 		BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1884 		sc->sc_if.if_flags |= IFF_OACTIVE;
1885 	} else {
1886 		sc->sc_if.if_flags &= ~IFF_OACTIVE;
1887 	}
1888 	if (sc->sc_if.if_flags & IFF_RUNNING) {
1889 		mutex_spin_enter(sc->sc_hwlock);
1890 		sc->sc_intmask |= XMTINT_0;
1891 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1892 		mutex_spin_exit(sc->sc_hwlock);
1893 	}
1894 	mutex_exit(sc->sc_lock);
1895 }
1896 #endif /* BCMETH_MPSAFETX */
1897 
1898 void
1899 bcmeth_soft_intr(void *arg)
1900 {
1901 	struct bcmeth_softc * const sc = arg;
1902 	struct ifnet * const ifp = &sc->sc_if;
1903 	uint32_t intmask = 0;
1904 
1905 	mutex_enter(sc->sc_lock);
1906 
1907 	u_int soft_flags = atomic_swap_uint(&sc->sc_soft_flags, 0);
1908 
1909 	BCMETH_EVCNT_INCR(sc->sc_ev_soft_intr);
1910 
1911 	if ((soft_flags & SOFT_TXINTR)
1912 	    || bcmeth_txq_active_p(sc, &sc->sc_txq)) {
1913 		/*
1914 		 * Let's do what we came here for.  Consume transmitted
1915 		 * packets off the the transmit ring.
1916 		 */
1917 		if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1918 		    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1919 			BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1920 			ifp->if_flags |= IFF_OACTIVE;
1921 		} else {
1922 			ifp->if_flags &= ~IFF_OACTIVE;
1923 		}
1924 		intmask |= XMTINT_0;
1925 	}
1926 
1927 	if (soft_flags & SOFT_RXINTR) {
1928 		/*
1929 		 * Let's consume
1930 		 */
1931 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
1932 		    sc->sc_rxq.rxq_threshold / 4)) {
1933 			/*
1934 			 * We've consumed a quarter of the ring and still have
1935 			 * more to do.  Refill the ring.
1936 			 */
1937 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
1938 		}
1939 		intmask |= RCVINT;
1940 	}
1941 
1942 	if (ifp->if_flags & IFF_RUNNING) {
1943 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
1944 		mutex_spin_enter(sc->sc_hwlock);
1945 		sc->sc_intmask |= intmask;
1946 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1947 		mutex_spin_exit(sc->sc_hwlock);
1948 	}
1949 
1950 	mutex_exit(sc->sc_lock);
1951 }
1952 
1953 void
1954 bcmeth_worker(struct work *wk, void *arg)
1955 {
1956 	struct bcmeth_softc * const sc = arg;
1957 	struct ifnet * const ifp = &sc->sc_if;
1958 	uint32_t intmask = 0;
1959 
1960 	mutex_enter(sc->sc_lock);
1961 
1962 	BCMETH_EVCNT_INCR(sc->sc_ev_work);
1963 
1964 	uint32_t work_flags = atomic_swap_32(&sc->sc_work_flags, 0);
1965 	if (work_flags & WORK_REINIT) {
1966 		int s = splnet();
1967 		sc->sc_soft_flags = 0;
1968 		bcmeth_ifinit(ifp);
1969 		splx(s);
1970 		work_flags &= ~WORK_RXUNDERFLOW;
1971 	}
1972 
1973 	if (work_flags & WORK_RXUNDERFLOW) {
1974 		struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
1975 		size_t threshold = 5 * rxq->rxq_threshold / 4;
1976 		if (threshold >= rxq->rxq_last - rxq->rxq_first) {
1977 			threshold = rxq->rxq_last - rxq->rxq_first - 1;
1978 		} else {
1979 			intmask |= RCVDESCUF;
1980 		}
1981 		aprint_normal_dev(sc->sc_dev,
1982 		    "increasing receive buffers from %zu to %zu\n",
1983 		    rxq->rxq_threshold, threshold);
1984 		rxq->rxq_threshold = threshold;
1985 	}
1986 
1987 	if (work_flags & WORK_RXINTR) {
1988 		/*
1989 		 * Let's consume
1990 		 */
1991 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
1992 		    sc->sc_rxq.rxq_threshold / 4)) {
1993 			/*
1994 			 * We've consumed a quarter of the ring and still have
1995 			 * more to do.  Refill the ring.
1996 			 */
1997 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
1998 		}
1999 		intmask |= RCVINT;
2000 	}
2001 
2002 	if (ifp->if_flags & IFF_RUNNING) {
2003 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
2004 #if 0
2005 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
2006 		if (intstatus & RCVINT) {
2007 			bcmeth_write_4(sc, GMAC_INTSTATUS, RCVINT);
2008 			work_flags |= WORK_RXINTR;
2009 			continue;
2010 		}
2011 #endif
2012 		mutex_spin_enter(sc->sc_hwlock);
2013 		sc->sc_intmask |= intmask;
2014 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
2015 		mutex_spin_exit(sc->sc_hwlock);
2016 	}
2017 
2018 	mutex_exit(sc->sc_lock);
2019 }
2020