xref: /netbsd-src/sys/arch/arm/broadcom/bcm53xx_eth.c (revision 413d532bcc3f62d122e56d92e13ac64825a40baf)
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 1.26 2014/02/23 21:19:06 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/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_attach(ifp);
405 	ether_ifattach(ifp, sc->sc_enaddr);
406 
407 #ifdef BCMETH_COUNTERS
408 	evcnt_attach_dynamic(&sc->sc_ev_intr, EVCNT_TYPE_INTR,
409 	    NULL, xname, "intr");
410 	evcnt_attach_dynamic(&sc->sc_ev_soft_intr, EVCNT_TYPE_INTR,
411 	    NULL, xname, "soft intr");
412 	evcnt_attach_dynamic(&sc->sc_ev_work, EVCNT_TYPE_MISC,
413 	    NULL, xname, "work items");
414 	evcnt_attach_dynamic(&sc->sc_ev_tx_stall, EVCNT_TYPE_MISC,
415 	    NULL, xname, "tx stalls");
416 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_lo, EVCNT_TYPE_MISC,
417 	    NULL, xname, "rx badmagic lo");
418 	evcnt_attach_dynamic(&sc->sc_ev_rx_badmagic_hi, EVCNT_TYPE_MISC,
419 	    NULL, xname, "rx badmagic hi");
420 #endif
421 }
422 
423 static int
424 bcmeth_mediachange(struct ifnet *ifp)
425 {
426 	//struct bcmeth_softc * const sc = ifp->if_softc;
427 	return 0;
428 }
429 
430 static void
431 bcmeth_mediastatus(struct ifnet *ifp, struct ifmediareq *ifm)
432 {
433 	//struct bcmeth_softc * const sc = ifp->if_softc;
434 
435 	ifm->ifm_status = IFM_AVALID | IFM_ACTIVE;
436 	ifm->ifm_active = IFM_ETHER | IFM_FDX | IFM_1000_T;
437 }
438 
439 static uint64_t
440 bcmeth_macaddr_create(const uint8_t *enaddr)
441 {
442 	return (enaddr[3] << 0)			// UNIMAC_MAC_0
443 	    |  (enaddr[2] << 8)			// UNIMAC_MAC_0
444 	    |  (enaddr[1] << 16)		// UNIMAC_MAC_0
445 	    |  ((uint64_t)enaddr[0] << 24)	// UNIMAC_MAC_0
446 	    |  ((uint64_t)enaddr[5] << 32)	// UNIMAC_MAC_1
447 	    |  ((uint64_t)enaddr[4] << 40);	// UNIMAC_MAC_1
448 }
449 
450 static int
451 bcmeth_ifinit(struct ifnet *ifp)
452 {
453 	struct bcmeth_softc * const sc = ifp->if_softc;
454 	int error = 0;
455 
456 	sc->sc_maxfrm = max(ifp->if_mtu + 32, MCLBYTES);
457 	if (ifp->if_mtu > ETHERMTU_JUMBO)
458 		return error;
459 
460 	KASSERT(ifp->if_flags & IFF_UP);
461 
462 	/*
463 	 * Stop the interface
464 	 */
465 	bcmeth_ifstop(ifp, 0);
466 
467 	/*
468 	 * Reserve enough space at the front so that we can insert a maxsized
469 	 * link header and a VLAN tag.  Also make sure we have enough room for
470 	 * the rcvsts field as well.
471 	 */
472 	KASSERT(ALIGN(max_linkhdr) == max_linkhdr);
473 	KASSERTMSG(max_linkhdr > sizeof(struct ether_header), "%u > %zu",
474 	    max_linkhdr, sizeof(struct ether_header));
475 	sc->sc_rcvoffset = max_linkhdr + 4 - sizeof(struct ether_header);
476 	if (sc->sc_rcvoffset <= 4)
477 		sc->sc_rcvoffset += 4;
478 	KASSERT((sc->sc_rcvoffset & 3) == 2);
479 	KASSERT(sc->sc_rcvoffset <= __SHIFTOUT(RCVCTL_RCVOFFSET, RCVCTL_RCVOFFSET));
480 	KASSERT(sc->sc_rcvoffset >= 6);
481 
482 	/*
483 	 * If our frame size has changed (or it's our first time through)
484 	 * destroy the existing transmit mapcache.
485 	 */
486 	if (sc->sc_tx_mapcache != NULL
487 	    && sc->sc_maxfrm != sc->sc_tx_mapcache->dmc_maxmapsize) {
488 		bcmeth_mapcache_destroy(sc, sc->sc_tx_mapcache);
489 		sc->sc_tx_mapcache = NULL;
490 	}
491 
492 	if (sc->sc_tx_mapcache == NULL) {
493 		error = bcmeth_mapcache_create(sc, &sc->sc_tx_mapcache,
494 		    BCMETH_MAXTXMBUFS, sc->sc_maxfrm, BCMETH_NTXSEGS);
495 		if (error)
496 			return error;
497 	}
498 
499 	sc->sc_cmdcfg = NO_LENGTH_CHECK | PAUSE_IGNORE
500 	    | __SHIFTIN(ETH_SPEED_1000, ETH_SPEED)
501 	    | RX_ENA | TX_ENA;
502 
503 	if (ifp->if_flags & IFF_PROMISC) {
504 		sc->sc_cmdcfg |= PROMISC_EN;
505 	} else {
506 		sc->sc_cmdcfg &= ~PROMISC_EN;
507 	}
508 
509 	const uint8_t * const lladdr = CLLADDR(ifp->if_sadl);
510 	const uint64_t macstnaddr = bcmeth_macaddr_create(lladdr);
511 
512 	/*
513 	 * We make sure that a received Ethernet packet start on a non-word
514 	 * boundary so that the packet payload will be on a word boundary.
515 	 * So to check the destination address we keep around two words to
516 	 * quickly compare with.
517 	 */
518 #if __ARMEL__
519 	sc->sc_macaddr[0] = lladdr[0] | (lladdr[1] << 8);
520 	sc->sc_macaddr[1] = lladdr[2] | (lladdr[3] << 8)
521 	    | (lladdr[4] << 16) | (lladdr[5] << 24);
522 #else
523 	sc->sc_macaddr[0] = lladdr[1] | (lladdr[0] << 8);
524 	sc->sc_macaddr[1] = lladdr[5] | (lladdr[4] << 8)
525 	    | (lladdr[1] << 16) | (lladdr[2] << 24);
526 #endif
527 
528 	sc->sc_intmask = DESCPROTOERR|DATAERR|DESCERR;
529 
530 	/* 5. Load RCVADDR_LO with new pointer */
531 	bcmeth_rxq_reset(sc, &sc->sc_rxq);
532 
533 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
534 	    __SHIFTIN(sc->sc_rcvoffset, RCVCTL_RCVOFFSET)
535 	    | RCVCTL_PARITY_DIS
536 	    | RCVCTL_OFLOW_CONTINUE
537 	    | __SHIFTIN(3, RCVCTL_BURSTLEN));
538 
539 	/* 6. Load XMTADDR_LO with new pointer */
540 	bcmeth_txq_reset(sc, &sc->sc_txq);
541 
542 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl, XMTCTL_DMA_ACT_INDEX
543 	    | XMTCTL_PARITY_DIS
544 	    | __SHIFTIN(3, XMTCTL_BURSTLEN));
545 
546 	/* 7. Setup other UNIMAC registers */
547 	bcmeth_write_4(sc, UNIMAC_FRAME_LEN, sc->sc_maxfrm);
548 	bcmeth_write_4(sc, UNIMAC_MAC_0, (uint32_t)(macstnaddr >>  0));
549 	bcmeth_write_4(sc, UNIMAC_MAC_1, (uint32_t)(macstnaddr >> 32));
550 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, sc->sc_cmdcfg);
551 
552 	uint32_t devctl = bcmeth_read_4(sc, GMAC_DEVCONTROL);
553 	devctl |= RGMII_LINK_STATUS_SEL | NWAY_AUTO_POLL_EN | TXARB_STRICT_MODE;
554 	devctl &= ~FLOW_CTRL_MODE;
555 	devctl &= ~MIB_RD_RESET_EN;
556 	devctl &= ~RXQ_OVERFLOW_CTRL_SEL;
557 	devctl &= ~CPU_FLOW_CTRL_ON;
558 	bcmeth_write_4(sc, GMAC_DEVCONTROL, devctl);
559 
560 	/* Setup lazy receive (at most 1ms). */
561 	const struct cpu_softc * const cpu = curcpu()->ci_softc;
562 	sc->sc_rcvlazy =  __SHIFTIN(4, INTRCVLAZY_FRAMECOUNT)
563 	     | __SHIFTIN(cpu->cpu_clk.clk_apb / 1000, INTRCVLAZY_TIMEOUT);
564 	bcmeth_write_4(sc, GMAC_INTRCVLAZY, sc->sc_rcvlazy);
565 
566 	/* 11. Enable transmit queues in TQUEUE, and ensure that the transmit scheduling mode is correctly set in TCTRL. */
567 	sc->sc_intmask |= XMTINT_0|XMTUF;
568 	bcmeth_write_4(sc, sc->sc_txq.txq_reg_xmtctl,
569 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl) | XMTCTL_ENABLE);
570 
571 
572 	/* 12. Enable receive queues in RQUEUE, */
573 	sc->sc_intmask |= RCVINT|RCVDESCUF|RCVFIFOOF;
574 	bcmeth_write_4(sc, sc->sc_rxq.rxq_reg_rcvctl,
575 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl) | RCVCTL_ENABLE);
576 
577 	bcmeth_rxq_produce(sc, &sc->sc_rxq);	/* fill with rx buffers */
578 
579 #if 0
580 	aprint_normal_dev(sc->sc_dev,
581 	    "devctl=%#x ucmdcfg=%#x xmtctl=%#x rcvctl=%#x\n",
582 	    devctl, sc->sc_cmdcfg,
583 	    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtctl),
584 	    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvctl));
585 #endif
586 
587 	sc->sc_soft_flags = 0;
588 
589 	bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
590 
591 	ifp->if_flags |= IFF_RUNNING;
592 
593 	return error;
594 }
595 
596 static void
597 bcmeth_ifstop(struct ifnet *ifp, int disable)
598 {
599 	struct bcmeth_softc * const sc = ifp->if_softc;
600 	struct bcmeth_txqueue * const txq = &sc->sc_txq;
601 	struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
602 
603 	KASSERT(!cpu_intr_p());
604 
605 	sc->sc_soft_flags = 0;
606 	sc->sc_work_flags = 0;
607 
608 	/* Disable Rx processing */
609 	bcmeth_write_4(sc, rxq->rxq_reg_rcvctl,
610 	    bcmeth_read_4(sc, rxq->rxq_reg_rcvctl) & ~RCVCTL_ENABLE);
611 
612 	/* Disable Tx processing */
613 	bcmeth_write_4(sc, txq->txq_reg_xmtctl,
614 	    bcmeth_read_4(sc, txq->txq_reg_xmtctl) & ~XMTCTL_ENABLE);
615 
616 	/* Disable all interrupts */
617 	bcmeth_write_4(sc, GMAC_INTMASK, 0);
618 
619 	for (;;) {
620 		uint32_t tx0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
621 		uint32_t rx0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
622 		if (__SHIFTOUT(tx0, XMTSTATE) == XMTSTATE_DIS
623 		    && __SHIFTOUT(rx0, RCVSTATE) == RCVSTATE_DIS)
624 			break;
625 		delay(50);
626 	}
627 	/*
628 	 * Now reset the controller.
629 	 *
630 	 * 3. Set SW_RESET bit in UNIMAC_COMMAND_CONFIG register
631 	 * 4. Clear SW_RESET bit in UNIMAC_COMMAND_CONFIG register
632 	 */
633 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, SW_RESET);
634 	bcmeth_write_4(sc, GMAC_INTSTATUS, ~0);
635 	sc->sc_intmask = 0;
636 	ifp->if_flags &= ~IFF_RUNNING;
637 
638 	/*
639 	 * Let's consume any remaining transmitted packets.  And if we are
640 	 * disabling the interface, purge ourselves of any untransmitted
641 	 * packets.  But don't consume any received packets, just drop them.
642 	 * If we aren't disabling the interface, save the mbufs in the
643 	 * receive queue for reuse.
644 	 */
645 	bcmeth_rxq_purge(sc, &sc->sc_rxq, disable);
646 	bcmeth_txq_consume(sc, &sc->sc_txq);
647 	if (disable) {
648 		bcmeth_txq_purge(sc, &sc->sc_txq);
649 		IF_PURGE(&ifp->if_snd);
650 	}
651 
652 	bcmeth_write_4(sc, UNIMAC_COMMAND_CONFIG, 0);
653 }
654 
655 static void
656 bcmeth_ifwatchdog(struct ifnet *ifp)
657 {
658 }
659 
660 static int
661 bcmeth_ifioctl(struct ifnet *ifp, u_long cmd, void *data)
662 {
663 	struct bcmeth_softc *sc  = ifp->if_softc;
664 	struct ifreq * const ifr = data;
665 	const int s = splnet();
666 	int error;
667 
668 	switch (cmd) {
669 	case SIOCSIFMEDIA:
670 	case SIOCGIFMEDIA:
671 		error = ifmedia_ioctl(ifp, ifr, &sc->sc_media, cmd);
672 		break;
673 
674 	default:
675 		error = ether_ioctl(ifp, cmd, data);
676 		if (error != ENETRESET)
677 			break;
678 
679 		if (cmd == SIOCADDMULTI || cmd == SIOCDELMULTI) {
680 			error = 0;
681 			break;
682 		}
683 		error = bcmeth_ifinit(ifp);
684 		break;
685 	}
686 
687 	splx(s);
688 	return error;
689 }
690 
691 static void
692 bcmeth_rxq_desc_presync(
693 	struct bcmeth_softc *sc,
694 	struct bcmeth_rxqueue *rxq,
695 	struct gmac_rxdb *rxdb,
696 	size_t count)
697 {
698 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
699 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
700 	    BUS_DMASYNC_PREWRITE);
701 }
702 
703 static void
704 bcmeth_rxq_desc_postsync(
705 	struct bcmeth_softc *sc,
706 	struct bcmeth_rxqueue *rxq,
707 	struct gmac_rxdb *rxdb,
708 	size_t count)
709 {
710 	bus_dmamap_sync(sc->sc_dmat, rxq->rxq_descmap,
711 	    (rxdb - rxq->rxq_first) * sizeof(*rxdb), count * sizeof(*rxdb),
712 	    BUS_DMASYNC_POSTWRITE);
713 }
714 
715 static void
716 bcmeth_txq_desc_presync(
717 	struct bcmeth_softc *sc,
718 	struct bcmeth_txqueue *txq,
719 	struct gmac_txdb *txdb,
720 	size_t count)
721 {
722 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
723 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
724 	    BUS_DMASYNC_PREWRITE);
725 }
726 
727 static void
728 bcmeth_txq_desc_postsync(
729 	struct bcmeth_softc *sc,
730 	struct bcmeth_txqueue *txq,
731 	struct gmac_txdb *txdb,
732 	size_t count)
733 {
734 	bus_dmamap_sync(sc->sc_dmat, txq->txq_descmap,
735 	    (txdb - txq->txq_first) * sizeof(*txdb), count * sizeof(*txdb),
736 	    BUS_DMASYNC_POSTWRITE);
737 }
738 
739 static bus_dmamap_t
740 bcmeth_mapcache_get(
741 	struct bcmeth_softc *sc,
742 	struct bcmeth_mapcache *dmc)
743 {
744 	KASSERT(dmc->dmc_nmaps > 0);
745 	KASSERT(dmc->dmc_maps[dmc->dmc_nmaps-1] != NULL);
746 	return dmc->dmc_maps[--dmc->dmc_nmaps];
747 }
748 
749 static void
750 bcmeth_mapcache_put(
751 	struct bcmeth_softc *sc,
752 	struct bcmeth_mapcache *dmc,
753 	bus_dmamap_t map)
754 {
755 	KASSERT(map != NULL);
756 	KASSERT(dmc->dmc_nmaps < dmc->dmc_maxmaps);
757 	dmc->dmc_maps[dmc->dmc_nmaps++] = map;
758 }
759 
760 static void
761 bcmeth_mapcache_destroy(
762 	struct bcmeth_softc *sc,
763 	struct bcmeth_mapcache *dmc)
764 {
765 	const size_t dmc_size =
766 	    offsetof(struct bcmeth_mapcache, dmc_maps[dmc->dmc_maxmaps]);
767 
768 	for (u_int i = 0; i < dmc->dmc_maxmaps; i++) {
769 		bus_dmamap_destroy(sc->sc_dmat, dmc->dmc_maps[i]);
770 	}
771 	kmem_intr_free(dmc, dmc_size);
772 }
773 
774 static int
775 bcmeth_mapcache_create(
776 	struct bcmeth_softc *sc,
777 	struct bcmeth_mapcache **dmc_p,
778 	size_t maxmaps,
779 	size_t maxmapsize,
780 	size_t maxseg)
781 {
782 	const size_t dmc_size =
783 	    offsetof(struct bcmeth_mapcache, dmc_maps[maxmaps]);
784 	struct bcmeth_mapcache * const dmc =
785 		kmem_intr_zalloc(dmc_size, KM_NOSLEEP);
786 
787 	dmc->dmc_maxmaps = maxmaps;
788 	dmc->dmc_nmaps = maxmaps;
789 	dmc->dmc_maxmapsize = maxmapsize;
790 	dmc->dmc_maxseg = maxseg;
791 
792 	for (u_int i = 0; i < maxmaps; i++) {
793 		int error = bus_dmamap_create(sc->sc_dmat, dmc->dmc_maxmapsize,
794 		     dmc->dmc_maxseg, dmc->dmc_maxmapsize, 0,
795 		     BUS_DMA_WAITOK|BUS_DMA_ALLOCNOW, &dmc->dmc_maps[i]);
796 		if (error) {
797 			aprint_error_dev(sc->sc_dev,
798 			    "failed to creat dma map cache "
799 			    "entry %u of %zu: %d\n",
800 			    i, maxmaps, error);
801 			while (i-- > 0) {
802 				bus_dmamap_destroy(sc->sc_dmat,
803 				    dmc->dmc_maps[i]);
804 			}
805 			kmem_intr_free(dmc, dmc_size);
806 			return error;
807 		}
808 		KASSERT(dmc->dmc_maps[i] != NULL);
809 	}
810 
811 	*dmc_p = dmc;
812 
813 	return 0;
814 }
815 
816 #if 0
817 static void
818 bcmeth_dmamem_free(
819 	bus_dma_tag_t dmat,
820 	size_t map_size,
821 	bus_dma_segment_t *seg,
822 	bus_dmamap_t map,
823 	void *kvap)
824 {
825 	bus_dmamap_destroy(dmat, map);
826 	bus_dmamem_unmap(dmat, kvap, map_size);
827 	bus_dmamem_free(dmat, seg, 1);
828 }
829 #endif
830 
831 static int
832 bcmeth_dmamem_alloc(
833 	bus_dma_tag_t dmat,
834 	size_t map_size,
835 	bus_dma_segment_t *seg,
836 	bus_dmamap_t *map,
837 	void **kvap)
838 {
839 	int error;
840 	int nseg;
841 
842 	*kvap = NULL;
843 	*map = NULL;
844 
845 	error = bus_dmamem_alloc(dmat, map_size, 2*PAGE_SIZE, 0,
846 	   seg, 1, &nseg, 0);
847 	if (error)
848 		return error;
849 
850 	KASSERT(nseg == 1);
851 
852 	error = bus_dmamem_map(dmat, seg, nseg, map_size, (void **)kvap, 0);
853 	if (error == 0) {
854 		error = bus_dmamap_create(dmat, map_size, 1, map_size, 0, 0,
855 		    map);
856 		if (error == 0) {
857 			error = bus_dmamap_load(dmat, *map, *kvap, map_size,
858 			    NULL, 0);
859 			if (error == 0)
860 				return 0;
861 			bus_dmamap_destroy(dmat, *map);
862 			*map = NULL;
863 		}
864 		bus_dmamem_unmap(dmat, *kvap, map_size);
865 		*kvap = NULL;
866 	}
867 	bus_dmamem_free(dmat, seg, nseg);
868 	return 0;
869 }
870 
871 static struct mbuf *
872 bcmeth_rx_buf_alloc(
873 	struct bcmeth_softc *sc)
874 {
875 	struct mbuf *m = m_gethdr(M_DONTWAIT, MT_DATA);
876 	if (m == NULL) {
877 		printf("%s:%d: %s\n", __func__, __LINE__, "m_gethdr");
878 		return NULL;
879 	}
880 	MCLGET(m, M_DONTWAIT);
881 	if ((m->m_flags & M_EXT) == 0) {
882 		printf("%s:%d: %s\n", __func__, __LINE__, "MCLGET");
883 		m_freem(m);
884 		return NULL;
885 	}
886 	m->m_len = m->m_pkthdr.len = m->m_ext.ext_size;
887 
888 	bus_dmamap_t map = bcmeth_mapcache_get(sc, sc->sc_rx_mapcache);
889 	if (map == NULL) {
890 		printf("%s:%d: %s\n", __func__, __LINE__, "map get");
891 		m_freem(m);
892 		return NULL;
893 	}
894 	M_SETCTX(m, map);
895 	m->m_len = m->m_pkthdr.len = MCLBYTES;
896 	int error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
897 	    BUS_DMA_READ|BUS_DMA_NOWAIT);
898 	if (error) {
899 		aprint_error_dev(sc->sc_dev, "fail to load rx dmamap: %d\n",
900 		    error);
901 		M_SETCTX(m, NULL);
902 		m_freem(m);
903 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
904 		return NULL;
905 	}
906 	KASSERT(map->dm_mapsize == MCLBYTES);
907 #ifdef BCMETH_RCVMAGIC
908 	*mtod(m, uint32_t *) = htole32(BCMETH_RCVMAGIC);
909 	bus_dmamap_sync(sc->sc_dmat, map, 0, sizeof(uint32_t),
910 	    BUS_DMASYNC_PREREAD|BUS_DMASYNC_PREWRITE);
911 	bus_dmamap_sync(sc->sc_dmat, map, sizeof(uint32_t),
912 	    map->dm_mapsize - sizeof(uint32_t), BUS_DMASYNC_PREREAD);
913 #else
914 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
915 	    BUS_DMASYNC_PREREAD);
916 #endif
917 
918 	return m;
919 }
920 
921 static void
922 bcmeth_rx_map_unload(
923 	struct bcmeth_softc *sc,
924 	struct mbuf *m)
925 {
926 	KASSERT(m);
927 	for (; m != NULL; m = m->m_next) {
928 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
929 		KASSERT(map);
930 		KASSERT(map->dm_mapsize == MCLBYTES);
931 		bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_len,
932 		    BUS_DMASYNC_POSTREAD);
933 		bus_dmamap_unload(sc->sc_dmat, map);
934 		bcmeth_mapcache_put(sc, sc->sc_rx_mapcache, map);
935 		M_SETCTX(m, NULL);
936 	}
937 }
938 
939 static bool
940 bcmeth_rxq_produce(
941 	struct bcmeth_softc *sc,
942 	struct bcmeth_rxqueue *rxq)
943 {
944 	struct gmac_rxdb *producer = rxq->rxq_producer;
945 	bool produced = false;
946 
947 	while (rxq->rxq_inuse < rxq->rxq_threshold) {
948 		struct mbuf *m;
949 		IF_DEQUEUE(&sc->sc_rx_bufcache, m);
950 		if (m == NULL) {
951 			m = bcmeth_rx_buf_alloc(sc);
952 			if (m == NULL) {
953 				printf("%s: bcmeth_rx_buf_alloc failed\n", __func__);
954 				break;
955 			}
956 		}
957 		bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
958 		KASSERT(map);
959 
960 		producer->rxdb_buflen = htole32(MCLBYTES);
961 		producer->rxdb_addrlo = htole32(map->dm_segs[0].ds_addr);
962 		producer->rxdb_flags &= htole32(RXDB_FLAG_ET);
963 		*rxq->rxq_mtail = m;
964 		rxq->rxq_mtail = &m->m_next;
965 		m->m_len = MCLBYTES;
966 		m->m_next = NULL;
967 		rxq->rxq_inuse++;
968 		if (++producer == rxq->rxq_last) {
969 			membar_producer();
970 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
971 			    rxq->rxq_last - rxq->rxq_producer);
972 			producer = rxq->rxq_producer = rxq->rxq_first;
973 		}
974 		produced = true;
975 	}
976 	if (produced) {
977 		membar_producer();
978 		if (producer != rxq->rxq_producer) {
979 			bcmeth_rxq_desc_presync(sc, rxq, rxq->rxq_producer,
980 			    producer - rxq->rxq_producer);
981 			rxq->rxq_producer = producer;
982 		}
983 		bcmeth_write_4(sc, rxq->rxq_reg_rcvptr,
984 		    rxq->rxq_descmap->dm_segs[0].ds_addr
985 		    + ((uintptr_t)producer & RCVPTR));
986 	}
987 	return true;
988 }
989 
990 static void
991 bcmeth_rx_input(
992 	struct bcmeth_softc *sc,
993 	struct mbuf *m,
994 	uint32_t rxdb_flags)
995 {
996 	struct ifnet * const ifp = &sc->sc_if;
997 
998 	bcmeth_rx_map_unload(sc, m);
999 
1000 	m_adj(m, sc->sc_rcvoffset);
1001 
1002 	/*
1003 	 * If we are in promiscuous mode and this isn't a multicast, check the
1004 	 * destination address to make sure it matches our own.  If it doesn't,
1005 	 * mark the packet as being received promiscuously.
1006 	 */
1007 	if ((sc->sc_cmdcfg & PROMISC_EN)
1008 	    && (m->m_data[0] & 1) == 0
1009 	    && (*(uint16_t *)&m->m_data[0] != sc->sc_macaddr[0]
1010 		|| *(uint32_t *)&m->m_data[2] != sc->sc_macaddr[1])) {
1011 		m->m_flags |= M_PROMISC;
1012 	}
1013 	m->m_pkthdr.rcvif = ifp;
1014 
1015 	ifp->if_ipackets++;
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 	(*ifp->if_input)(ifp, m);
1024 	mutex_enter(sc->sc_lock);
1025 #else
1026 	int s = splnet();
1027 	bpf_mtap(ifp, m);
1028 	(*ifp->if_input)(ifp, m);
1029 	splx(s);
1030 #endif
1031 }
1032 
1033 static bool
1034 bcmeth_rxq_consume(
1035 	struct bcmeth_softc *sc,
1036 	struct bcmeth_rxqueue *rxq,
1037 	size_t atmost)
1038 {
1039 	struct ifnet * const ifp = &sc->sc_if;
1040 	struct gmac_rxdb *consumer = rxq->rxq_consumer;
1041 	size_t rxconsumed = 0;
1042 	bool didconsume = false;
1043 
1044 	while (atmost-- > 0) {
1045 		if (consumer == rxq->rxq_producer) {
1046 			KASSERT(rxq->rxq_inuse == 0);
1047 			break;
1048 		}
1049 
1050 		uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1051 		uint32_t currdscr = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1052 		if (consumer == rxq->rxq_first + currdscr) {
1053 			break;
1054 		}
1055 		bcmeth_rxq_desc_postsync(sc, rxq, consumer, 1);
1056 
1057 		/*
1058 		 * We own this packet again.  Copy the rxsts word from it.
1059 		 */
1060 		rxconsumed++;
1061 		didconsume = true;
1062 		uint32_t rxsts;
1063 		KASSERT(rxq->rxq_mhead != NULL);
1064 		bus_dmamap_t map = M_GETCTX(rxq->rxq_mhead, bus_dmamap_t);
1065 		bus_dmamap_sync(sc->sc_dmat, map, 0, arm_dcache_align,
1066 		    BUS_DMASYNC_POSTREAD);
1067 		memcpy(&rxsts, rxq->rxq_mhead->m_data, 4);
1068 		rxsts = le32toh(rxsts);
1069 #if 0
1070 		KASSERTMSG(rxsts != BCMETH_RCVMAGIC, "currdscr=%u consumer=%zd",
1071 		    currdscr, consumer - rxq->rxq_first);
1072 #endif
1073 
1074 		/*
1075 		 * Get the count of descriptors.  Fetch the correct number
1076 		 * of mbufs.
1077 		 */
1078 #ifdef BCMETH_RCVMAGIC
1079 		size_t desc_count = rxsts != BCMETH_RCVMAGIC ? __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1 : 1;
1080 #else
1081 		size_t desc_count = __SHIFTOUT(rxsts, RXSTS_DESC_COUNT) + 1;
1082 #endif
1083 		struct mbuf *m = rxq->rxq_mhead;
1084 		struct mbuf *m_last = m;
1085 		for (size_t i = 1; i < desc_count; i++) {
1086 			if (++consumer == rxq->rxq_last) {
1087 				consumer = rxq->rxq_first;
1088 			}
1089 			KASSERTMSG(consumer != rxq->rxq_first + currdscr,
1090 			    "i=%zu rxsts=%#x desc_count=%zu currdscr=%u consumer=%zd",
1091 			    i, rxsts, desc_count, currdscr,
1092 			    consumer - rxq->rxq_first);
1093 			m_last = m_last->m_next;
1094 		}
1095 
1096 		/*
1097 		 * Now remove it/them from the list of enqueued mbufs.
1098 		 */
1099 		if ((rxq->rxq_mhead = m_last->m_next) == NULL)
1100 			rxq->rxq_mtail = &rxq->rxq_mhead;
1101 		m_last->m_next = NULL;
1102 
1103 #ifdef BCMETH_RCVMAGIC
1104 		if (rxsts == BCMETH_RCVMAGIC) {
1105 			ifp->if_ierrors++;
1106 			if ((m->m_ext.ext_paddr >> 28) == 8) {
1107 				BCMETH_EVCNT_INCR(sc->sc_ev_rx_badmagic_lo);
1108 			} else {
1109 				BCMETH_EVCNT_INCR( sc->sc_ev_rx_badmagic_hi);
1110 			}
1111 			IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1112 		} else
1113 #endif /* BCMETH_RCVMAGIC */
1114 		if (rxsts & (RXSTS_CRC_ERROR|RXSTS_OVERSIZED|RXSTS_PKT_OVERFLOW)) {
1115 			aprint_error_dev(sc->sc_dev, "[%zu]: count=%zu rxsts=%#x\n",
1116 			    consumer - rxq->rxq_first, desc_count, rxsts);
1117 			/*
1118 			 * We encountered an error, take the mbufs and add them
1119 			 * to the rx bufcache so we can quickly reuse them.
1120 			 */
1121 			ifp->if_ierrors++;
1122 			do {
1123 				struct mbuf *m0 = m->m_next;
1124 				m->m_next = NULL;
1125 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1126 				m = m0;
1127 			} while (m);
1128 		} else {
1129 			uint32_t framelen = __SHIFTOUT(rxsts, RXSTS_FRAMELEN);
1130 			framelen += sc->sc_rcvoffset;
1131 			m->m_pkthdr.len = framelen;
1132 			if (desc_count == 1) {
1133 				KASSERT(framelen <= MCLBYTES);
1134 				m->m_len = framelen;
1135 			} else {
1136 				m_last->m_len = framelen & (MCLBYTES - 1);
1137 			}
1138 
1139 #ifdef BCMETH_MPSAFE
1140 			/*
1141 			 * Wrap at the last entry!
1142 			 */
1143 			if (++consumer == rxq->rxq_last) {
1144 				KASSERT(consumer[-1].rxdb_flags & htole32(RXDB_FLAG_ET));
1145 				rxq->rxq_consumer = rxq->rxq_first;
1146 			} else {
1147 				rxq->rxq_consumer = consumer;
1148 			}
1149 			rxq->rxq_inuse -= rxconsumed;
1150 #endif /* BCMETH_MPSAFE */
1151 
1152 			/*
1153 			 * Receive the packet (which releases our lock)
1154 			 */
1155 			bcmeth_rx_input(sc, m, rxsts);
1156 
1157 #ifdef BCMETH_MPSAFE
1158 			/*
1159 			 * Since we had to give up our lock, we need to
1160 			 * refresh these.
1161 			 */
1162 			consumer = rxq->rxq_consumer;
1163 			rxconsumed = 0;
1164 			continue;
1165 #endif /* BCMETH_MPSAFE */
1166 		}
1167 
1168 		/*
1169 		 * Wrap at the last entry!
1170 		 */
1171 		if (++consumer == rxq->rxq_last) {
1172 			KASSERT(consumer[-1].rxdb_flags & htole32(RXDB_FLAG_ET));
1173 			consumer = rxq->rxq_first;
1174 		}
1175 	}
1176 
1177 	/*
1178 	 * Update queue info.
1179 	 */
1180 	rxq->rxq_consumer = consumer;
1181 	rxq->rxq_inuse -= rxconsumed;
1182 
1183 	/*
1184 	 * Did we consume anything?
1185 	 */
1186 	return didconsume;
1187 }
1188 
1189 static void
1190 bcmeth_rxq_purge(
1191 	struct bcmeth_softc *sc,
1192 	struct bcmeth_rxqueue *rxq,
1193 	bool discard)
1194 {
1195 	struct mbuf *m;
1196 
1197 	if ((m = rxq->rxq_mhead) != NULL) {
1198 		if (discard) {
1199 			bcmeth_rx_map_unload(sc, m);
1200 			m_freem(m);
1201 		} else {
1202 			while (m != NULL) {
1203 				struct mbuf *m0 = m->m_next;
1204 				m->m_next = NULL;
1205 				IF_ENQUEUE(&sc->sc_rx_bufcache, m);
1206 				m = m0;
1207 			}
1208 		}
1209 
1210 	}
1211 
1212 	rxq->rxq_mhead = NULL;
1213 	rxq->rxq_mtail = &rxq->rxq_mhead;
1214 	rxq->rxq_inuse = 0;
1215 }
1216 
1217 static void
1218 bcmeth_rxq_reset(
1219 	struct bcmeth_softc *sc,
1220 	struct bcmeth_rxqueue *rxq)
1221 {
1222 	/*
1223 	 * sync all the descriptors
1224 	 */
1225 	bcmeth_rxq_desc_postsync(sc, rxq, rxq->rxq_first,
1226 	    rxq->rxq_last - rxq->rxq_first);
1227 
1228 	/*
1229 	 * Make sure we own all descriptors in the ring.
1230 	 */
1231 	struct gmac_rxdb *rxdb;
1232 	for (rxdb = rxq->rxq_first; rxdb < rxq->rxq_last - 1; rxdb++) {
1233 		rxdb->rxdb_flags = htole32(RXDB_FLAG_IC);
1234 	}
1235 
1236 	/*
1237 	 * Last descriptor has the wrap flag.
1238 	 */
1239 	rxdb->rxdb_flags = htole32(RXDB_FLAG_ET|RXDB_FLAG_IC);
1240 
1241 	/*
1242 	 * Reset the producer consumer indexes.
1243 	 */
1244 	rxq->rxq_consumer = rxq->rxq_first;
1245 	rxq->rxq_producer = rxq->rxq_first;
1246 	rxq->rxq_inuse = 0;
1247 	if (rxq->rxq_threshold < BCMETH_MINRXMBUFS)
1248 		rxq->rxq_threshold = BCMETH_MINRXMBUFS;
1249 
1250 	sc->sc_intmask |= RCVINT|RCVFIFOOF|RCVDESCUF;
1251 
1252 	/*
1253 	 * Restart the receiver at the first descriptor
1254 	 */
1255 	bcmeth_write_4(sc, rxq->rxq_reg_rcvaddrlo,
1256 	    rxq->rxq_descmap->dm_segs[0].ds_addr);
1257 }
1258 
1259 static int
1260 bcmeth_rxq_attach(
1261 	struct bcmeth_softc *sc,
1262 	struct bcmeth_rxqueue *rxq,
1263 	u_int qno)
1264 {
1265 	size_t desc_count = BCMETH_RINGSIZE / sizeof(rxq->rxq_first[0]);
1266 	int error;
1267 	void *descs;
1268 
1269 	KASSERT(desc_count == 256 || desc_count == 512);
1270 
1271 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1272 	   &rxq->rxq_descmap_seg, &rxq->rxq_descmap, &descs);
1273 	if (error)
1274 		return error;
1275 
1276 	memset(descs, 0, BCMETH_RINGSIZE);
1277 	rxq->rxq_first = descs;
1278 	rxq->rxq_last = rxq->rxq_first + desc_count;
1279 	rxq->rxq_consumer = descs;
1280 	rxq->rxq_producer = descs;
1281 
1282 	bcmeth_rxq_purge(sc, rxq, true);
1283 	bcmeth_rxq_reset(sc, rxq);
1284 
1285 	rxq->rxq_reg_rcvaddrlo = GMAC_RCVADDR_LOW;
1286 	rxq->rxq_reg_rcvctl = GMAC_RCVCONTROL;
1287 	rxq->rxq_reg_rcvptr = GMAC_RCVPTR;
1288 	rxq->rxq_reg_rcvsts0 = GMAC_RCVSTATUS0;
1289 	rxq->rxq_reg_rcvsts1 = GMAC_RCVSTATUS1;
1290 
1291 	return 0;
1292 }
1293 
1294 static bool
1295 bcmeth_txq_active_p(
1296 	struct bcmeth_softc * const sc,
1297 	struct bcmeth_txqueue *txq)
1298 {
1299 	return !IF_IS_EMPTY(&txq->txq_mbufs);
1300 }
1301 
1302 static bool
1303 bcmeth_txq_fillable_p(
1304 	struct bcmeth_softc * const sc,
1305 	struct bcmeth_txqueue *txq)
1306 {
1307 	return txq->txq_free >= txq->txq_threshold;
1308 }
1309 
1310 static int
1311 bcmeth_txq_attach(
1312 	struct bcmeth_softc *sc,
1313 	struct bcmeth_txqueue *txq,
1314 	u_int qno)
1315 {
1316 	size_t desc_count = BCMETH_RINGSIZE / sizeof(txq->txq_first[0]);
1317 	int error;
1318 	void *descs;
1319 
1320 	KASSERT(desc_count == 256 || desc_count == 512);
1321 
1322 	error = bcmeth_dmamem_alloc(sc->sc_dmat, BCMETH_RINGSIZE,
1323 	   &txq->txq_descmap_seg, &txq->txq_descmap, &descs);
1324 	if (error)
1325 		return error;
1326 
1327 	memset(descs, 0, BCMETH_RINGSIZE);
1328 	txq->txq_first = descs;
1329 	txq->txq_last = txq->txq_first + desc_count;
1330 	txq->txq_consumer = descs;
1331 	txq->txq_producer = descs;
1332 
1333 	IFQ_SET_MAXLEN(&txq->txq_mbufs, BCMETH_MAXTXMBUFS);
1334 
1335 	txq->txq_reg_xmtaddrlo = GMAC_XMTADDR_LOW;
1336 	txq->txq_reg_xmtctl = GMAC_XMTCONTROL;
1337 	txq->txq_reg_xmtptr = GMAC_XMTPTR;
1338 	txq->txq_reg_xmtsts0 = GMAC_XMTSTATUS0;
1339 	txq->txq_reg_xmtsts1 = GMAC_XMTSTATUS1;
1340 
1341 	bcmeth_txq_reset(sc, txq);
1342 
1343 	return 0;
1344 }
1345 
1346 static int
1347 bcmeth_txq_map_load(
1348 	struct bcmeth_softc *sc,
1349 	struct bcmeth_txqueue *txq,
1350 	struct mbuf *m)
1351 {
1352 	bus_dmamap_t map;
1353 	int error;
1354 
1355 	map = M_GETCTX(m, bus_dmamap_t);
1356 	if (map != NULL)
1357 		return 0;
1358 
1359 	map = bcmeth_mapcache_get(sc, sc->sc_tx_mapcache);
1360 	if (map == NULL)
1361 		return ENOMEM;
1362 
1363 	error = bus_dmamap_load_mbuf(sc->sc_dmat, map, m,
1364 	    BUS_DMA_WRITE | BUS_DMA_NOWAIT);
1365 	if (error)
1366 		return error;
1367 
1368 	bus_dmamap_sync(sc->sc_dmat, map, 0, m->m_pkthdr.len,
1369 	    BUS_DMASYNC_PREWRITE);
1370 	M_SETCTX(m, map);
1371 	return 0;
1372 }
1373 
1374 static void
1375 bcmeth_txq_map_unload(
1376 	struct bcmeth_softc *sc,
1377 	struct bcmeth_txqueue *txq,
1378 	struct mbuf *m)
1379 {
1380 	KASSERT(m);
1381 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1382 	bus_dmamap_sync(sc->sc_dmat, map, 0, map->dm_mapsize,
1383 	    BUS_DMASYNC_POSTWRITE);
1384 	bus_dmamap_unload(sc->sc_dmat, map);
1385 	bcmeth_mapcache_put(sc, sc->sc_tx_mapcache, map);
1386 }
1387 
1388 static bool
1389 bcmeth_txq_produce(
1390 	struct bcmeth_softc *sc,
1391 	struct bcmeth_txqueue *txq,
1392 	struct mbuf *m)
1393 {
1394 	bus_dmamap_t map = M_GETCTX(m, bus_dmamap_t);
1395 
1396 	if (map->dm_nsegs > txq->txq_free)
1397 		return false;
1398 
1399 	/*
1400 	 * TCP Offload flag must be set in the first descriptor.
1401 	 */
1402 	struct gmac_txdb *producer = txq->txq_producer;
1403 	uint32_t first_flags = TXDB_FLAG_SF;
1404 	uint32_t last_flags = TXDB_FLAG_EF;
1405 
1406 	/*
1407 	 * If we've produced enough descriptors without consuming any
1408 	 * we need to ask for an interrupt to reclaim some.
1409 	 */
1410 	txq->txq_lastintr += map->dm_nsegs;
1411 	if (txq->txq_lastintr >= txq->txq_threshold
1412 	    || txq->txq_mbufs.ifq_len + 1 == txq->txq_mbufs.ifq_maxlen) {
1413 		txq->txq_lastintr = 0;
1414 		last_flags |= TXDB_FLAG_IC;
1415 	}
1416 
1417 	KASSERT(producer != txq->txq_last);
1418 
1419 	struct gmac_txdb *start = producer;
1420 	size_t count = map->dm_nsegs;
1421 	producer->txdb_flags |= htole32(first_flags);
1422 	producer->txdb_addrlo = htole32(map->dm_segs[0].ds_addr);
1423 	producer->txdb_buflen = htole32(map->dm_segs[0].ds_len);
1424 	for (u_int i = 1; i < map->dm_nsegs; i++) {
1425 #if 0
1426 		printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1427 		    le32toh(producer->txdb_flags),
1428 		    le32toh(producer->txdb_buflen),
1429 		    le32toh(producer->txdb_addrlo),
1430 		    le32toh(producer->txdb_addrhi));
1431 #endif
1432 		if (__predict_false(++producer == txq->txq_last)) {
1433 			bcmeth_txq_desc_presync(sc, txq, start,
1434 			    txq->txq_last - start);
1435 			count -= txq->txq_last - start;
1436 			producer = txq->txq_first;
1437 			start = txq->txq_first;
1438 		}
1439 		producer->txdb_addrlo = htole32(map->dm_segs[i].ds_addr);
1440 		producer->txdb_buflen = htole32(map->dm_segs[i].ds_len);
1441 	}
1442 	producer->txdb_flags |= htole32(last_flags);
1443 #if 0
1444 	printf("[%zu]: %#x/%#x/%#x/%#x\n", producer - txq->txq_first,
1445 	    le32toh(producer->txdb_flags), le32toh(producer->txdb_buflen),
1446 	    le32toh(producer->txdb_addrlo), le32toh(producer->txdb_addrhi));
1447 #endif
1448 	if (count)
1449 		bcmeth_txq_desc_presync(sc, txq, start, count);
1450 
1451 	/*
1452 	 * Reduce free count by the number of segments we consumed.
1453 	 */
1454 	txq->txq_free -= map->dm_nsegs;
1455 	KASSERT(map->dm_nsegs == 1 || txq->txq_producer != producer);
1456 	KASSERT(map->dm_nsegs == 1 || (txq->txq_producer->txdb_flags & htole32(TXDB_FLAG_EF)) == 0);
1457 	KASSERT(producer->txdb_flags & htole32(TXDB_FLAG_EF));
1458 
1459 #if 0
1460 	printf("%s: mbuf %p: produced a %u byte packet in %u segments (%zd..%zd)\n",
1461 	    __func__, m, m->m_pkthdr.len, map->dm_nsegs,
1462 	    txq->txq_producer - txq->txq_first, producer - txq->txq_first);
1463 #endif
1464 
1465 	if (producer + 1 == txq->txq_last)
1466 		txq->txq_producer = txq->txq_first;
1467 	else
1468 		txq->txq_producer = producer + 1;
1469 	IF_ENQUEUE(&txq->txq_mbufs, m);
1470 
1471 	/*
1472 	 * Let the transmitter know there's more to do
1473 	 */
1474 	bcmeth_write_4(sc, txq->txq_reg_xmtptr,
1475 	    txq->txq_descmap->dm_segs[0].ds_addr
1476 	    + ((uintptr_t)txq->txq_producer & XMT_LASTDSCR));
1477 
1478 	return true;
1479 }
1480 
1481 static struct mbuf *
1482 bcmeth_copy_packet(struct mbuf *m)
1483 {
1484 	struct mbuf *mext = NULL;
1485 	size_t misalignment = 0;
1486 	size_t hlen = 0;
1487 
1488 	for (mext = m; mext != NULL; mext = mext->m_next) {
1489 		if (mext->m_flags & M_EXT) {
1490 			misalignment = mtod(mext, vaddr_t) & arm_dcache_align;
1491 			break;
1492 		}
1493 		hlen += m->m_len;
1494 	}
1495 
1496 	struct mbuf *n = m->m_next;
1497 	if (m != mext && hlen + misalignment <= MHLEN && false) {
1498 		KASSERT(m->m_pktdat <= m->m_data && m->m_data <= &m->m_pktdat[MHLEN - m->m_len]);
1499 		size_t oldoff = m->m_data - m->m_pktdat;
1500 		size_t off;
1501 		if (mext == NULL) {
1502 			off = (oldoff + hlen > MHLEN) ? 0 : oldoff;
1503 		} else {
1504 			off = MHLEN - (hlen + misalignment);
1505 		}
1506 		KASSERT(off + hlen + misalignment <= MHLEN);
1507 		if (((oldoff ^ off) & arm_dcache_align) != 0 || off < oldoff) {
1508 			memmove(&m->m_pktdat[off], m->m_data, m->m_len);
1509 			m->m_data = &m->m_pktdat[off];
1510 		}
1511 		m_copydata(n, 0, hlen - m->m_len, &m->m_data[m->m_len]);
1512 		m->m_len = hlen;
1513 		m->m_next = mext;
1514 		while (n != mext) {
1515 			n = m_free(n);
1516 		}
1517 		return m;
1518 	}
1519 
1520 	struct mbuf *m0 = m_gethdr(M_DONTWAIT, m->m_type);
1521 	if (m0 == NULL) {
1522 		return NULL;
1523 	}
1524 	M_COPY_PKTHDR(m0, m);
1525 	MCLAIM(m0, m->m_owner);
1526 	if (m0->m_pkthdr.len > MHLEN) {
1527 		MCLGET(m0, M_DONTWAIT);
1528 		if ((m0->m_flags & M_EXT) == 0) {
1529 			m_freem(m0);
1530 			return NULL;
1531 		}
1532 	}
1533 	m0->m_len = m->m_pkthdr.len;
1534 	m_copydata(m, 0, m0->m_len, mtod(m0, void *));
1535 	m_freem(m);
1536 	return m0;
1537 }
1538 
1539 static bool
1540 bcmeth_txq_enqueue(
1541 	struct bcmeth_softc *sc,
1542 	struct bcmeth_txqueue *txq)
1543 {
1544 	for (;;) {
1545 		if (IF_QFULL(&txq->txq_mbufs))
1546 			return false;
1547 		struct mbuf *m = txq->txq_next;
1548 		if (m == NULL) {
1549 			int s = splnet();
1550 			IF_DEQUEUE(&sc->sc_if.if_snd, m);
1551 			splx(s);
1552 			if (m == NULL)
1553 				return true;
1554 			M_SETCTX(m, NULL);
1555 		} else {
1556 			txq->txq_next = NULL;
1557 		}
1558 		/*
1559 		 * If LINK2 is set and this packet uses multiple mbufs,
1560 		 * consolidate it into a single mbuf.
1561 		 */
1562 		if (m->m_next != NULL && (sc->sc_if.if_flags & IFF_LINK2)) {
1563 			struct mbuf *m0 = bcmeth_copy_packet(m);
1564 			if (m0 == NULL) {
1565 				txq->txq_next = m;
1566 				return true;
1567 			}
1568 			m = m0;
1569 		}
1570 		int error = bcmeth_txq_map_load(sc, txq, m);
1571 		if (error) {
1572 			aprint_error_dev(sc->sc_dev,
1573 			    "discarded packet due to "
1574 			    "dmamap load failure: %d\n", error);
1575 			m_freem(m);
1576 			continue;
1577 		}
1578 		KASSERT(txq->txq_next == NULL);
1579 		if (!bcmeth_txq_produce(sc, txq, m)) {
1580 			txq->txq_next = m;
1581 			return false;
1582 		}
1583 		KASSERT(txq->txq_next == NULL);
1584 	}
1585 }
1586 
1587 static bool
1588 bcmeth_txq_consume(
1589 	struct bcmeth_softc *sc,
1590 	struct bcmeth_txqueue *txq)
1591 {
1592 	struct ifnet * const ifp = &sc->sc_if;
1593 	struct gmac_txdb *consumer = txq->txq_consumer;
1594 	size_t txfree = 0;
1595 
1596 #if 0
1597 	printf("%s: entry: free=%zu\n", __func__, txq->txq_free);
1598 #endif
1599 
1600 	for (;;) {
1601 		if (consumer == txq->txq_producer) {
1602 			txq->txq_consumer = consumer;
1603 			txq->txq_free += txfree;
1604 			txq->txq_lastintr -= min(txq->txq_lastintr, txfree);
1605 #if 0
1606 			printf("%s: empty: freed %zu descriptors going from %zu to %zu\n",
1607 			    __func__, txfree, txq->txq_free - txfree, txq->txq_free);
1608 #endif
1609 			KASSERT(txq->txq_lastintr == 0);
1610 			KASSERT(txq->txq_free == txq->txq_last - txq->txq_first - 1);
1611 			return true;
1612 		}
1613 		bcmeth_txq_desc_postsync(sc, txq, consumer, 1);
1614 		uint32_t s0 = bcmeth_read_4(sc, txq->txq_reg_xmtsts0);
1615 		if (consumer == txq->txq_first + __SHIFTOUT(s0, XMT_CURRDSCR)) {
1616 			txq->txq_consumer = consumer;
1617 			txq->txq_free += txfree;
1618 			txq->txq_lastintr -= min(txq->txq_lastintr, txfree);
1619 #if 0
1620 			printf("%s: freed %zu descriptors\n",
1621 			    __func__, txfree);
1622 #endif
1623 			return bcmeth_txq_fillable_p(sc, txq);
1624 		}
1625 
1626 		/*
1627 		 * If this is the last descriptor in the chain, get the
1628 		 * mbuf, free its dmamap, and free the mbuf chain itself.
1629 		 */
1630 		const uint32_t txdb_flags = le32toh(consumer->txdb_flags);
1631 		if (txdb_flags & TXDB_FLAG_EF) {
1632 			struct mbuf *m;
1633 
1634 			IF_DEQUEUE(&txq->txq_mbufs, m);
1635 			KASSERT(m);
1636 			bcmeth_txq_map_unload(sc, txq, m);
1637 #if 0
1638 			printf("%s: mbuf %p: consumed a %u byte packet\n",
1639 			    __func__, m, m->m_pkthdr.len);
1640 #endif
1641 			bpf_mtap(ifp, m);
1642 			ifp->if_opackets++;
1643 			ifp->if_obytes += m->m_pkthdr.len;
1644 			if (m->m_flags & M_MCAST)
1645 				ifp->if_omcasts++;
1646 			m_freem(m);
1647 		}
1648 
1649 		/*
1650 		 * We own this packet again.  Clear all flags except wrap.
1651 		 */
1652 		txfree++;
1653 
1654 		/*
1655 		 * Wrap at the last entry!
1656 		 */
1657 		if (txdb_flags & TXDB_FLAG_ET) {
1658 			consumer->txdb_flags = htole32(TXDB_FLAG_ET);
1659 			KASSERT(consumer + 1 == txq->txq_last);
1660 			consumer = txq->txq_first;
1661 		} else {
1662 			consumer->txdb_flags = 0;
1663 			consumer++;
1664 			KASSERT(consumer < txq->txq_last);
1665 		}
1666 	}
1667 }
1668 
1669 static void
1670 bcmeth_txq_purge(
1671 	struct bcmeth_softc *sc,
1672 	struct bcmeth_txqueue *txq)
1673 {
1674 	struct mbuf *m;
1675 	KASSERT((bcmeth_read_4(sc, UNIMAC_COMMAND_CONFIG) & TX_ENA) == 0);
1676 
1677 	for (;;) {
1678 		IF_DEQUEUE(&txq->txq_mbufs, m);
1679 		if (m == NULL)
1680 			break;
1681 		bcmeth_txq_map_unload(sc, txq, m);
1682 		m_freem(m);
1683 	}
1684 	if ((m = txq->txq_next) != NULL) {
1685 		txq->txq_next = NULL;
1686 		bcmeth_txq_map_unload(sc, txq, m);
1687 		m_freem(m);
1688 	}
1689 }
1690 
1691 static void
1692 bcmeth_txq_reset(
1693 	struct bcmeth_softc *sc,
1694 	struct bcmeth_txqueue *txq)
1695 {
1696 	/*
1697 	 * sync all the descriptors
1698 	 */
1699 	bcmeth_txq_desc_postsync(sc, txq, txq->txq_first,
1700 	    txq->txq_last - txq->txq_first);
1701 
1702 	/*
1703 	 * Make sure we own all descriptors in the ring.
1704 	 */
1705 	struct gmac_txdb *txdb;
1706 	for (txdb = txq->txq_first; txdb < txq->txq_last - 1; txdb++) {
1707 		txdb->txdb_flags = 0;
1708 	}
1709 
1710 	/*
1711 	 * Last descriptor has the wrap flag.
1712 	 */
1713 	txdb->txdb_flags = htole32(TXDB_FLAG_ET);
1714 
1715 	/*
1716 	 * Reset the producer consumer indexes.
1717 	 */
1718 	txq->txq_consumer = txq->txq_first;
1719 	txq->txq_producer = txq->txq_first;
1720 	txq->txq_free = txq->txq_last - txq->txq_first - 1;
1721 	txq->txq_threshold = txq->txq_free / 2;
1722 	txq->txq_lastintr = 0;
1723 
1724 	/*
1725 	 * What do we want to get interrupted on?
1726 	 */
1727 	sc->sc_intmask |= XMTINT_0 | XMTUF;
1728 
1729 	/*
1730 	 * Restart the transmiter at the first descriptor
1731 	 */
1732 	bcmeth_write_4(sc, txq->txq_reg_xmtaddrlo,
1733 	    txq->txq_descmap->dm_segs->ds_addr);
1734 }
1735 
1736 static void
1737 bcmeth_ifstart(struct ifnet *ifp)
1738 {
1739 	struct bcmeth_softc * const sc = ifp->if_softc;
1740 
1741 	if (__predict_false((ifp->if_flags & IFF_RUNNING) == 0)) {
1742 		return;
1743 	}
1744 
1745 #ifdef BCMETH_MPSAFETX
1746 	if (cpu_intr_p()) {
1747 #endif
1748 		atomic_or_uint(&sc->sc_soft_flags, SOFT_TXINTR);
1749 		softint_schedule(sc->sc_soft_ih);
1750 #ifdef BCMETH_MPSAFETX
1751 	} else {
1752 		/*
1753 		 * Either we are in a softintr thread already or some other
1754 		 * thread so just borrow it to do the send and save ourselves
1755 		 * the overhead of a fast soft int.
1756 		 */
1757 		bcmeth_soft_txintr(sc);
1758 	}
1759 #endif
1760 }
1761 
1762 int
1763 bcmeth_intr(void *arg)
1764 {
1765 	struct bcmeth_softc * const sc = arg;
1766 	uint32_t soft_flags = 0;
1767 	uint32_t work_flags = 0;
1768 	int rv = 0;
1769 
1770 	mutex_enter(sc->sc_hwlock);
1771 
1772 	uint32_t intmask = sc->sc_intmask;
1773 	BCMETH_EVCNT_INCR(sc->sc_ev_intr);
1774 
1775 	for (;;) {
1776 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
1777 		intstatus &= intmask;
1778 		bcmeth_write_4(sc, GMAC_INTSTATUS, intstatus);	/* write 1 to clear */
1779 		if (intstatus == 0) {
1780 			break;
1781 		}
1782 #if 0
1783 		aprint_normal_dev(sc->sc_dev, "%s: intstatus=%#x intmask=%#x\n",
1784 		    __func__, intstatus, bcmeth_read_4(sc, GMAC_INTMASK));
1785 #endif
1786 		if (intstatus & RCVINT) {
1787 			struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
1788 			intmask &= ~RCVINT;
1789 
1790 			uint32_t rcvsts0 = bcmeth_read_4(sc, rxq->rxq_reg_rcvsts0);
1791 			uint32_t descs = __SHIFTOUT(rcvsts0, RCV_CURRDSCR);
1792 			if (descs < rxq->rxq_consumer - rxq->rxq_first) {
1793 				/*
1794 				 * We wrapped at the end so count how far
1795 				 * we are from the end.
1796 				 */
1797 				descs += rxq->rxq_last - rxq->rxq_consumer;
1798 			} else {
1799 				descs -= rxq->rxq_consumer - rxq->rxq_first;
1800 			}
1801 			/*
1802 			 * If we "timedout" we can't be hogging so use
1803 			 * softints.  If we exceeded then we might hogging
1804 			 * so let the workqueue deal with them.
1805 			 */
1806 			const uint32_t framecount = __SHIFTOUT(sc->sc_rcvlazy, INTRCVLAZY_FRAMECOUNT);
1807 			if (descs < framecount
1808 			    || (curcpu()->ci_curlwp->l_flag & LW_IDLE)) {
1809 				soft_flags |= SOFT_RXINTR;
1810 			} else {
1811 				work_flags |= WORK_RXINTR;
1812 			}
1813 		}
1814 
1815 		if (intstatus & XMTINT_0) {
1816 			intmask &= ~XMTINT_0;
1817 			soft_flags |= SOFT_TXINTR;
1818 		}
1819 
1820 		if (intstatus & RCVDESCUF) {
1821 			intmask &= ~RCVDESCUF;
1822 			work_flags |= WORK_RXUNDERFLOW;
1823 		}
1824 
1825 		intstatus &= intmask;
1826 		if (intstatus) {
1827 			aprint_error_dev(sc->sc_dev,
1828 			    "intr: intstatus=%#x\n", intstatus);
1829 			aprint_error_dev(sc->sc_dev,
1830 			    "rcvbase=%p/%#lx rcvptr=%#x rcvsts=%#x/%#x\n",
1831 			    sc->sc_rxq.rxq_first,
1832 			    sc->sc_rxq.rxq_descmap->dm_segs[0].ds_addr,
1833 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvptr),
1834 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts0),
1835 			    bcmeth_read_4(sc, sc->sc_rxq.rxq_reg_rcvsts1));
1836 			aprint_error_dev(sc->sc_dev,
1837 			    "xmtbase=%p/%#lx xmtptr=%#x xmtsts=%#x/%#x\n",
1838 			    sc->sc_txq.txq_first,
1839 			    sc->sc_txq.txq_descmap->dm_segs[0].ds_addr,
1840 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtptr),
1841 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts0),
1842 			    bcmeth_read_4(sc, sc->sc_txq.txq_reg_xmtsts1));
1843 			intmask &= ~intstatus;
1844 			work_flags |= WORK_REINIT;
1845 			break;
1846 		}
1847 	}
1848 
1849 	if (intmask != sc->sc_intmask) {
1850 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1851 	}
1852 
1853 	if (work_flags) {
1854 		if (sc->sc_work_flags == 0) {
1855 			workqueue_enqueue(sc->sc_workq, &sc->sc_work, NULL);
1856 		}
1857 		atomic_or_32(&sc->sc_work_flags, work_flags);
1858 		rv = 1;
1859 	}
1860 
1861 	if (soft_flags) {
1862 		if (sc->sc_soft_flags == 0) {
1863 			softint_schedule(sc->sc_soft_ih);
1864 		}
1865 		atomic_or_32(&sc->sc_soft_flags, soft_flags);
1866 		rv = 1;
1867 	}
1868 
1869 	mutex_exit(sc->sc_hwlock);
1870 
1871 	return rv;
1872 }
1873 
1874 #ifdef BCMETH_MPSAFETX
1875 void
1876 bcmeth_soft_txintr(struct bcmeth_softc *sc)
1877 {
1878 	mutex_enter(sc->sc_lock);
1879 	/*
1880 	 * Let's do what we came here for.  Consume transmitted
1881 	 * packets off the the transmit ring.
1882 	 */
1883 	if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1884 	    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1885 		BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1886 		sc->sc_if.if_flags |= IFF_OACTIVE;
1887 	} else {
1888 		sc->sc_if.if_flags &= ~IFF_OACTIVE;
1889 	}
1890 	if (sc->sc_if.if_flags & IFF_RUNNING) {
1891 		mutex_spin_enter(sc->sc_hwlock);
1892 		sc->sc_intmask |= XMTINT_0;
1893 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1894 		mutex_spin_exit(sc->sc_hwlock);
1895 	}
1896 	mutex_exit(sc->sc_lock);
1897 }
1898 #endif /* BCMETH_MPSAFETX */
1899 
1900 void
1901 bcmeth_soft_intr(void *arg)
1902 {
1903 	struct bcmeth_softc * const sc = arg;
1904 	struct ifnet * const ifp = &sc->sc_if;
1905 	uint32_t intmask = 0;
1906 
1907 	mutex_enter(sc->sc_lock);
1908 
1909 	u_int soft_flags = atomic_swap_uint(&sc->sc_soft_flags, 0);
1910 
1911 	BCMETH_EVCNT_INCR(sc->sc_ev_soft_intr);
1912 
1913 	if ((soft_flags & SOFT_TXINTR)
1914 	    || bcmeth_txq_active_p(sc, &sc->sc_txq)) {
1915 		/*
1916 		 * Let's do what we came here for.  Consume transmitted
1917 		 * packets off the the transmit ring.
1918 		 */
1919 		if (!bcmeth_txq_consume(sc, &sc->sc_txq)
1920 		    || !bcmeth_txq_enqueue(sc, &sc->sc_txq)) {
1921 			BCMETH_EVCNT_INCR(sc->sc_ev_tx_stall);
1922 			ifp->if_flags |= IFF_OACTIVE;
1923 		} else {
1924 			ifp->if_flags &= ~IFF_OACTIVE;
1925 		}
1926 		intmask |= XMTINT_0;
1927 	}
1928 
1929 	if (soft_flags & SOFT_RXINTR) {
1930 		/*
1931 		 * Let's consume
1932 		 */
1933 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
1934 		    sc->sc_rxq.rxq_threshold / 4)) {
1935 			/*
1936 			 * We've consumed a quarter of the ring and still have
1937 			 * more to do.  Refill the ring.
1938 			 */
1939 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
1940 		}
1941 		intmask |= RCVINT;
1942 	}
1943 
1944 	if (ifp->if_flags & IFF_RUNNING) {
1945 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
1946 		mutex_spin_enter(sc->sc_hwlock);
1947 		sc->sc_intmask |= intmask;
1948 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
1949 		mutex_spin_exit(sc->sc_hwlock);
1950 	}
1951 
1952 	mutex_exit(sc->sc_lock);
1953 }
1954 
1955 void
1956 bcmeth_worker(struct work *wk, void *arg)
1957 {
1958 	struct bcmeth_softc * const sc = arg;
1959 	struct ifnet * const ifp = &sc->sc_if;
1960 	uint32_t intmask = 0;
1961 
1962 	mutex_enter(sc->sc_lock);
1963 
1964 	BCMETH_EVCNT_INCR(sc->sc_ev_work);
1965 
1966 	uint32_t work_flags = atomic_swap_32(&sc->sc_work_flags, 0);
1967 	if (work_flags & WORK_REINIT) {
1968 		int s = splnet();
1969 		sc->sc_soft_flags = 0;
1970 		bcmeth_ifinit(ifp);
1971 		splx(s);
1972 		work_flags &= ~WORK_RXUNDERFLOW;
1973 	}
1974 
1975 	if (work_flags & WORK_RXUNDERFLOW) {
1976 		struct bcmeth_rxqueue * const rxq = &sc->sc_rxq;
1977 		size_t threshold = 5 * rxq->rxq_threshold / 4;
1978 		if (threshold >= rxq->rxq_last - rxq->rxq_first) {
1979 			threshold = rxq->rxq_last - rxq->rxq_first - 1;
1980 		} else {
1981 			intmask |= RCVDESCUF;
1982 		}
1983 		aprint_normal_dev(sc->sc_dev,
1984 		    "increasing receive buffers from %zu to %zu\n",
1985 		    rxq->rxq_threshold, threshold);
1986 		rxq->rxq_threshold = threshold;
1987 	}
1988 
1989 	if (work_flags & WORK_RXINTR) {
1990 		/*
1991 		 * Let's consume
1992 		 */
1993 		while (bcmeth_rxq_consume(sc, &sc->sc_rxq,
1994 		    sc->sc_rxq.rxq_threshold / 4)) {
1995 			/*
1996 			 * We've consumed a quarter of the ring and still have
1997 			 * more to do.  Refill the ring.
1998 			 */
1999 			bcmeth_rxq_produce(sc, &sc->sc_rxq);
2000 		}
2001 		intmask |= RCVINT;
2002 	}
2003 
2004 	if (ifp->if_flags & IFF_RUNNING) {
2005 		bcmeth_rxq_produce(sc, &sc->sc_rxq);
2006 #if 0
2007 		uint32_t intstatus = bcmeth_read_4(sc, GMAC_INTSTATUS);
2008 		if (intstatus & RCVINT) {
2009 			bcmeth_write_4(sc, GMAC_INTSTATUS, RCVINT);
2010 			work_flags |= WORK_RXINTR;
2011 			continue;
2012 		}
2013 #endif
2014 		mutex_spin_enter(sc->sc_hwlock);
2015 		sc->sc_intmask |= intmask;
2016 		bcmeth_write_4(sc, GMAC_INTMASK, sc->sc_intmask);
2017 		mutex_spin_exit(sc->sc_hwlock);
2018 	}
2019 
2020 	mutex_exit(sc->sc_lock);
2021 }
2022