xref: /titanic_51/usr/src/uts/common/io/dmfe/dmfe_main.c (revision 53a7b6b6763f5865522a76e5e887390a8f4777d7)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 
27 #include <sys/types.h>
28 #include <sys/sunddi.h>
29 #include "dmfe_impl.h"
30 
31 /*
32  * This is the string displayed by modinfo, etc.
33  */
34 static char dmfe_ident[] = "Davicom DM9102 Ethernet";
35 
36 
37 /*
38  * NOTES:
39  *
40  * #defines:
41  *
42  *	DMFE_PCI_RNUMBER is the register-set number to use for the operating
43  *	registers.  On an OBP-based machine, regset 0 refers to CONFIG space,
44  *	regset 1 will be the operating registers in I/O space, and regset 2
45  *	will be the operating registers in MEMORY space (preferred).  If an
46  *	expansion ROM is fitted, it may appear as a further register set.
47  *
48  *	DMFE_SLOP defines the amount by which the chip may read beyond
49  *	the end of a buffer or descriptor, apparently 6-8 dwords :(
50  *	We have to make sure this doesn't cause it to access unallocated
51  *	or unmapped memory.
52  *
53  *	DMFE_BUF_SIZE must be at least (ETHERMAX + ETHERFCSL + DMFE_SLOP)
54  *	rounded up to a multiple of 4.  Here we choose a power of two for
55  *	speed & simplicity at the cost of a bit more memory.
56  *
57  *	However, the buffer length field in the TX/RX descriptors is only
58  *	eleven bits, so even though we allocate DMFE_BUF_SIZE (2048) bytes
59  *	per buffer, we tell the chip that they're only DMFE_BUF_SIZE_1
60  *	(2000) bytes each.
61  *
62  *	DMFE_DMA_MODE defines the mode (STREAMING/CONSISTENT) used for
63  *	the data buffers.  The descriptors are always set up in CONSISTENT
64  *	mode.
65  *
66  *	DMFE_HEADROOM defines how much space we'll leave in allocated
67  *	mblks before the first valid data byte.  This should be chosen
68  *	to be 2 modulo 4, so that once the ethernet header (14 bytes)
69  *	has been stripped off, the packet data will be 4-byte aligned.
70  *	The remaining space can be used by upstream modules to prepend
71  *	any headers required.
72  *
73  * Patchable globals:
74  *
75  *	dmfe_bus_modes: the bus mode bits to be put into CSR0.
76  *		Setting READ_MULTIPLE in this register seems to cause
77  *		the chip to generate a READ LINE command with a parity
78  *		error!  Don't do it!
79  *
80  *	dmfe_setup_desc1: the value to be put into descriptor word 1
81  *		when sending a SETUP packet.
82  *
83  *		Setting TX_LAST_DESC in desc1 in a setup packet seems
84  *		to make the chip spontaneously reset internally - it
85  *		attempts to give back the setup packet descriptor by
86  *		writing to PCI address 00000000 - which may or may not
87  *		get a MASTER ABORT - after which most of its registers
88  *		seem to have either default values or garbage!
89  *
90  *		TX_FIRST_DESC doesn't seem to have the same effect but
91  *		it isn't needed on a setup packet so we'll leave it out
92  *		too, just in case it has some other wierd side-effect.
93  *
94  *		The default hardware packet filtering mode is now
95  *		HASH_AND_PERFECT (imperfect filtering of multicast
96  *		packets and perfect filtering of unicast packets).
97  *		If this is found not to work reliably, setting the
98  *		TX_FILTER_TYPE1 bit will cause a switchover to using
99  *		HASH_ONLY mode (imperfect filtering of *all* packets).
100  *		Software will then perform the additional filtering
101  *		as required.
102  */
103 
104 #define	DMFE_PCI_RNUMBER	2
105 #define	DMFE_SLOP		(8*sizeof (uint32_t))
106 #define	DMFE_BUF_SIZE		2048
107 #define	DMFE_BUF_SIZE_1		2000
108 #define	DMFE_DMA_MODE		DDI_DMA_STREAMING
109 #define	DMFE_HEADROOM		34
110 
111 static uint32_t dmfe_bus_modes = TX_POLL_INTVL | CACHE_ALIGN;
112 static uint32_t dmfe_setup_desc1 = TX_SETUP_PACKET | SETUPBUF_SIZE |
113 					TX_FILTER_TYPE0;
114 
115 /*
116  * Some tunable parameters ...
117  *	Number of RX/TX ring entries (128/128)
118  *	Minimum number of TX ring slots to keep free (1)
119  *	Low-water mark at which to try to reclaim TX ring slots (1)
120  *	How often to take a TX-done interrupt (twice per ring cycle)
121  *	Whether to reclaim TX ring entries on a TX-done interrupt (no)
122  */
123 
124 #define	DMFE_TX_DESC		128	/* Should be a multiple of 4 <= 256 */
125 #define	DMFE_RX_DESC		128	/* Should be a multiple of 4 <= 256 */
126 
127 static uint32_t dmfe_rx_desc = DMFE_RX_DESC;
128 static uint32_t dmfe_tx_desc = DMFE_TX_DESC;
129 static uint32_t dmfe_tx_min_free = 1;
130 static uint32_t dmfe_tx_reclaim_level = 1;
131 static uint32_t dmfe_tx_int_factor = (DMFE_TX_DESC / 2) - 1;
132 static boolean_t dmfe_reclaim_on_done = B_FALSE;
133 
134 /*
135  * Time-related parameters:
136  *
137  *	We use a cyclic to provide a periodic callback; this is then used
138  * 	to check for TX-stall and poll the link status register.
139  *
140  *	DMFE_TICK is the interval between cyclic callbacks, in microseconds.
141  *
142  *	TX_STALL_TIME_100 is the timeout in microseconds between passing
143  *	a packet to the chip for transmission and seeing that it's gone,
144  *	when running at 100Mb/s.  If we haven't reclaimed at least one
145  *	descriptor in this time we assume the transmitter has stalled
146  *	and reset the chip.
147  *
148  *	TX_STALL_TIME_10 is the equivalent timeout when running at 10Mb/s.
149  *
150  *	LINK_POLL_TIME is the interval between checks on the link state
151  *	when nothing appears to have happened (this is in addition to the
152  *	case where we think we've detected a link change, and serves as a
153  *	backup in case the quick link check doesn't work properly).
154  *
155  * Patchable globals:
156  *
157  *	dmfe_tick_us:		DMFE_TICK
158  *	dmfe_tx100_stall_us:	TX_STALL_TIME_100
159  *	dmfe_tx10_stall_us:	TX_STALL_TIME_10
160  *	dmfe_link_poll_us:	LINK_POLL_TIME
161  *
162  * These are then used in _init() to calculate:
163  *
164  *	stall_100_tix[]: number of consecutive cyclic callbacks without a
165  *			 reclaim before the TX process is considered stalled,
166  *			 when running at 100Mb/s.  The elements are indexed
167  *			 by transmit-engine-state.
168  *	stall_10_tix[]:	 number of consecutive cyclic callbacks without a
169  *			 reclaim before the TX process is considered stalled,
170  *			 when running at 10Mb/s.  The elements are indexed
171  *			 by transmit-engine-state.
172  *	factotum_tix:	 number of consecutive cyclic callbacks before waking
173  *			 up the factotum even though there doesn't appear to
174  *			 be anything for it to do
175  */
176 
177 #define	DMFE_TICK		25000		/* microseconds		*/
178 #define	TX_STALL_TIME_100	50000		/* microseconds		*/
179 #define	TX_STALL_TIME_10	200000		/* microseconds		*/
180 #define	LINK_POLL_TIME		5000000		/* microseconds		*/
181 
182 static uint32_t dmfe_tick_us = DMFE_TICK;
183 static uint32_t dmfe_tx100_stall_us = TX_STALL_TIME_100;
184 static uint32_t dmfe_tx10_stall_us = TX_STALL_TIME_10;
185 static uint32_t dmfe_link_poll_us = LINK_POLL_TIME;
186 
187 /*
188  * Calculated from above in _init()
189  */
190 
191 static uint32_t stall_100_tix[TX_PROCESS_MAX_STATE+1];
192 static uint32_t stall_10_tix[TX_PROCESS_MAX_STATE+1];
193 static uint32_t factotum_tix;
194 static uint32_t factotum_fast_tix;
195 static uint32_t factotum_start_tix;
196 
197 /*
198  * Property names
199  */
200 static char localmac_propname[] = "local-mac-address";
201 static char opmode_propname[] = "opmode-reg-value";
202 static char debug_propname[] = "dmfe-debug-flags";
203 
204 static int		dmfe_m_start(void *);
205 static void		dmfe_m_stop(void *);
206 static int		dmfe_m_promisc(void *, boolean_t);
207 static int		dmfe_m_multicst(void *, boolean_t, const uint8_t *);
208 static int		dmfe_m_unicst(void *, const uint8_t *);
209 static void		dmfe_m_ioctl(void *, queue_t *, mblk_t *);
210 static mblk_t		*dmfe_m_tx(void *, mblk_t *);
211 static int 		dmfe_m_stat(void *, uint_t, uint64_t *);
212 
213 static mac_callbacks_t dmfe_m_callbacks = {
214 	(MC_IOCTL),
215 	dmfe_m_stat,
216 	dmfe_m_start,
217 	dmfe_m_stop,
218 	dmfe_m_promisc,
219 	dmfe_m_multicst,
220 	dmfe_m_unicst,
221 	dmfe_m_tx,
222 	dmfe_m_ioctl,
223 	NULL,
224 };
225 
226 
227 /*
228  * Describes the chip's DMA engine
229  */
230 static ddi_dma_attr_t dma_attr = {
231 	DMA_ATTR_V0,		/* dma_attr version */
232 	0,			/* dma_attr_addr_lo */
233 	(uint32_t)0xFFFFFFFF,	/* dma_attr_addr_hi */
234 	0x0FFFFFF,		/* dma_attr_count_max */
235 	0x20,			/* dma_attr_align */
236 	0x7F,			/* dma_attr_burstsizes */
237 	1,			/* dma_attr_minxfer */
238 	(uint32_t)0xFFFFFFFF,	/* dma_attr_maxxfer */
239 	(uint32_t)0xFFFFFFFF,	/* dma_attr_seg */
240 	1,			/* dma_attr_sgllen */
241 	1,			/* dma_attr_granular */
242 	0			/* dma_attr_flags */
243 };
244 
245 /*
246  * DMA access attributes for registers and descriptors
247  */
248 static ddi_device_acc_attr_t dmfe_reg_accattr = {
249 	DDI_DEVICE_ATTR_V0,
250 	DDI_STRUCTURE_LE_ACC,
251 	DDI_STRICTORDER_ACC
252 };
253 
254 /*
255  * DMA access attributes for data: NOT to be byte swapped.
256  */
257 static ddi_device_acc_attr_t dmfe_data_accattr = {
258 	DDI_DEVICE_ATTR_V0,
259 	DDI_NEVERSWAP_ACC,
260 	DDI_STRICTORDER_ACC
261 };
262 
263 static uchar_t dmfe_broadcast_addr[ETHERADDRL] = {
264 	0xff, 0xff, 0xff, 0xff, 0xff, 0xff
265 };
266 
267 
268 /*
269  * ========== Lowest-level chip register & ring access routines ==========
270  */
271 
272 /*
273  * I/O register get/put routines
274  */
275 uint32_t
276 dmfe_chip_get32(dmfe_t *dmfep, off_t offset)
277 {
278 	uint32_t *addr;
279 
280 	addr = (void *)(dmfep->io_reg + offset);
281 	return (ddi_get32(dmfep->io_handle, addr));
282 }
283 
284 void
285 dmfe_chip_put32(dmfe_t *dmfep, off_t offset, uint32_t value)
286 {
287 	uint32_t *addr;
288 
289 	addr = (void *)(dmfep->io_reg + offset);
290 	ddi_put32(dmfep->io_handle, addr, value);
291 }
292 
293 /*
294  * TX/RX ring get/put routines
295  */
296 static uint32_t
297 dmfe_ring_get32(dma_area_t *dma_p, uint_t index, uint_t offset)
298 {
299 	uint32_t *addr;
300 
301 	addr = (void *)dma_p->mem_va;
302 	return (ddi_get32(dma_p->acc_hdl, addr + index*DESC_SIZE + offset));
303 }
304 
305 static void
306 dmfe_ring_put32(dma_area_t *dma_p, uint_t index, uint_t offset, uint32_t value)
307 {
308 	uint32_t *addr;
309 
310 	addr = (void *)dma_p->mem_va;
311 	ddi_put32(dma_p->acc_hdl, addr + index*DESC_SIZE + offset, value);
312 }
313 
314 /*
315  * Setup buffer get/put routines
316  */
317 static uint32_t
318 dmfe_setup_get32(dma_area_t *dma_p, uint_t index)
319 {
320 	uint32_t *addr;
321 
322 	addr = (void *)dma_p->setup_va;
323 	return (ddi_get32(dma_p->acc_hdl, addr + index));
324 }
325 
326 static void
327 dmfe_setup_put32(dma_area_t *dma_p, uint_t index, uint32_t value)
328 {
329 	uint32_t *addr;
330 
331 	addr = (void *)dma_p->setup_va;
332 	ddi_put32(dma_p->acc_hdl, addr + index, value);
333 }
334 
335 
336 /*
337  * ========== Low-level chip & ring buffer manipulation ==========
338  */
339 
340 #define	DMFE_DBG	DMFE_DBG_REGS	/* debug flag for this code	*/
341 
342 /*
343  * dmfe_set_opmode() -- function to set operating mode
344  */
345 static void
346 dmfe_set_opmode(dmfe_t *dmfep)
347 {
348 	DMFE_DEBUG(("dmfe_set_opmode: opmode 0x%x", dmfep->opmode));
349 
350 	ASSERT(mutex_owned(dmfep->oplock));
351 
352 	dmfe_chip_put32(dmfep, OPN_MODE_REG, dmfep->opmode);
353 	drv_usecwait(10);
354 }
355 
356 /*
357  * dmfe_stop_chip() -- stop all chip processing & optionally reset the h/w
358  */
359 static void
360 dmfe_stop_chip(dmfe_t *dmfep, enum chip_state newstate)
361 {
362 	ASSERT(mutex_owned(dmfep->oplock));
363 
364 	/*
365 	 * Stop the chip:
366 	 *	disable all interrupts
367 	 *	stop TX/RX processes
368 	 *	clear the status bits for TX/RX stopped
369 	 * If required, reset the chip
370 	 * Record the new state
371 	 */
372 	dmfe_chip_put32(dmfep, INT_MASK_REG, 0);
373 	dmfep->opmode &= ~(START_TRANSMIT | START_RECEIVE);
374 	dmfe_set_opmode(dmfep);
375 	dmfe_chip_put32(dmfep, STATUS_REG, TX_STOPPED_INT | RX_STOPPED_INT);
376 
377 	switch (newstate) {
378 	default:
379 		ASSERT(!"can't get here");
380 		return;
381 
382 	case CHIP_STOPPED:
383 	case CHIP_ERROR:
384 		break;
385 
386 	case CHIP_RESET:
387 		dmfe_chip_put32(dmfep, BUS_MODE_REG, SW_RESET);
388 		drv_usecwait(10);
389 		dmfe_chip_put32(dmfep, BUS_MODE_REG, 0);
390 		drv_usecwait(10);
391 		dmfe_chip_put32(dmfep, BUS_MODE_REG, dmfe_bus_modes);
392 		break;
393 	}
394 
395 	dmfep->chip_state = newstate;
396 }
397 
398 /*
399  * Initialize transmit and receive descriptor rings, and
400  * set the chip to point to the first entry in each ring
401  */
402 static void
403 dmfe_init_rings(dmfe_t *dmfep)
404 {
405 	dma_area_t *descp;
406 	uint32_t pstart;
407 	uint32_t pnext;
408 	uint32_t pbuff;
409 	uint32_t desc1;
410 	int i;
411 
412 	/*
413 	 * You need all the locks in order to rewrite the descriptor rings
414 	 */
415 	ASSERT(mutex_owned(dmfep->oplock));
416 	ASSERT(mutex_owned(dmfep->rxlock));
417 	ASSERT(mutex_owned(dmfep->txlock));
418 
419 	/*
420 	 * Program the RX ring entries
421 	 */
422 	descp = &dmfep->rx_desc;
423 	pstart = descp->mem_dvma;
424 	pnext = pstart + sizeof (struct rx_desc_type);
425 	pbuff = dmfep->rx_buff.mem_dvma;
426 	desc1 = RX_CHAINING | DMFE_BUF_SIZE_1;
427 
428 	for (i = 0; i < dmfep->rx.n_desc; ++i) {
429 		dmfe_ring_put32(descp, i, RD_NEXT, pnext);
430 		dmfe_ring_put32(descp, i, BUFFER1, pbuff);
431 		dmfe_ring_put32(descp, i, DESC1, desc1);
432 		dmfe_ring_put32(descp, i, DESC0, RX_OWN);
433 
434 		pnext += sizeof (struct rx_desc_type);
435 		pbuff += DMFE_BUF_SIZE;
436 	}
437 
438 	/*
439 	 * Fix up last entry & sync
440 	 */
441 	dmfe_ring_put32(descp, --i, RD_NEXT, pstart);
442 	DMA_SYNC(descp, DDI_DMA_SYNC_FORDEV);
443 	dmfep->rx.next_free = 0;
444 
445 	/*
446 	 * Set the base address of the RX descriptor list in CSR3
447 	 */
448 	DMFE_DEBUG(("RX descriptor VA: $%p (DVMA $%x)",
449 	    descp->mem_va, descp->mem_dvma));
450 	dmfe_chip_put32(dmfep, RX_BASE_ADDR_REG, descp->mem_dvma);
451 
452 	/*
453 	 * Program the TX ring entries
454 	 */
455 	descp = &dmfep->tx_desc;
456 	pstart = descp->mem_dvma;
457 	pnext = pstart + sizeof (struct tx_desc_type);
458 	pbuff = dmfep->tx_buff.mem_dvma;
459 	desc1 = TX_CHAINING;
460 
461 	for (i = 0; i < dmfep->tx.n_desc; ++i) {
462 		dmfe_ring_put32(descp, i, TD_NEXT, pnext);
463 		dmfe_ring_put32(descp, i, BUFFER1, pbuff);
464 		dmfe_ring_put32(descp, i, DESC1, desc1);
465 		dmfe_ring_put32(descp, i, DESC0, 0);
466 
467 		pnext += sizeof (struct tx_desc_type);
468 		pbuff += DMFE_BUF_SIZE;
469 	}
470 
471 	/*
472 	 * Fix up last entry & sync
473 	 */
474 	dmfe_ring_put32(descp, --i, TD_NEXT, pstart);
475 	DMA_SYNC(descp, DDI_DMA_SYNC_FORDEV);
476 	dmfep->tx.n_free = dmfep->tx.n_desc;
477 	dmfep->tx.next_free = dmfep->tx.next_busy = 0;
478 
479 	/*
480 	 * Set the base address of the TX descrptor list in CSR4
481 	 */
482 	DMFE_DEBUG(("TX descriptor VA: $%p (DVMA $%x)",
483 	    descp->mem_va, descp->mem_dvma));
484 	dmfe_chip_put32(dmfep, TX_BASE_ADDR_REG, descp->mem_dvma);
485 }
486 
487 /*
488  * dmfe_start_chip() -- start the chip transmitting and/or receiving
489  */
490 static void
491 dmfe_start_chip(dmfe_t *dmfep, int mode)
492 {
493 	ASSERT(mutex_owned(dmfep->oplock));
494 
495 	dmfep->opmode |= mode;
496 	dmfe_set_opmode(dmfep);
497 
498 	dmfe_chip_put32(dmfep, W_J_TIMER_REG, 0);
499 	/*
500 	 * Enable VLAN length mode (allows packets to be 4 bytes Longer).
501 	 */
502 	dmfe_chip_put32(dmfep, W_J_TIMER_REG, VLAN_ENABLE);
503 
504 	/*
505 	 * Clear any pending process-stopped interrupts
506 	 */
507 	dmfe_chip_put32(dmfep, STATUS_REG, TX_STOPPED_INT | RX_STOPPED_INT);
508 	dmfep->chip_state = mode & START_RECEIVE ? CHIP_TX_RX :
509 	    mode & START_TRANSMIT ? CHIP_TX_ONLY : CHIP_STOPPED;
510 }
511 
512 /*
513  * dmfe_enable_interrupts() -- enable our favourite set of interrupts.
514  *
515  * Normal interrupts:
516  *	We always enable:
517  *		RX_PKTDONE_INT		(packet received)
518  *		TX_PKTDONE_INT		(TX complete)
519  *	We never enable:
520  *		TX_ALLDONE_INT		(next TX buffer not ready)
521  *
522  * Abnormal interrupts:
523  *	We always enable:
524  *		RX_STOPPED_INT
525  *		TX_STOPPED_INT
526  *		SYSTEM_ERR_INT
527  *		RX_UNAVAIL_INT
528  *	We never enable:
529  *		RX_EARLY_INT
530  *		RX_WATCHDOG_INT
531  *		TX_JABBER_INT
532  *		TX_EARLY_INT
533  *		TX_UNDERFLOW_INT
534  *		GP_TIMER_INT		(not valid in -9 chips)
535  *		LINK_STATUS_INT		(not valid in -9 chips)
536  */
537 static void
538 dmfe_enable_interrupts(dmfe_t *dmfep)
539 {
540 	ASSERT(mutex_owned(dmfep->oplock));
541 
542 	/*
543 	 * Put 'the standard set of interrupts' in the interrupt mask register
544 	 */
545 	dmfep->imask =	RX_PKTDONE_INT | TX_PKTDONE_INT |
546 	    RX_STOPPED_INT | TX_STOPPED_INT | RX_UNAVAIL_INT | SYSTEM_ERR_INT;
547 
548 	dmfe_chip_put32(dmfep, INT_MASK_REG,
549 	    NORMAL_SUMMARY_INT | ABNORMAL_SUMMARY_INT | dmfep->imask);
550 	dmfep->chip_state = CHIP_RUNNING;
551 
552 	DMFE_DEBUG(("dmfe_enable_interrupts: imask 0x%x", dmfep->imask));
553 }
554 
555 #undef	DMFE_DBG
556 
557 
558 /*
559  * ========== RX side routines ==========
560  */
561 
562 #define	DMFE_DBG	DMFE_DBG_RECV	/* debug flag for this code	*/
563 
564 /*
565  * Function to update receive statistics on various errors
566  */
567 static void
568 dmfe_update_rx_stats(dmfe_t *dmfep, uint32_t desc0)
569 {
570 	ASSERT(mutex_owned(dmfep->rxlock));
571 
572 	/*
573 	 * The error summary bit and the error bits that it summarises
574 	 * are only valid if this is the last fragment.  Therefore, a
575 	 * fragment only contributes to the error statistics if both
576 	 * the last-fragment and error summary bits are set.
577 	 */
578 	if (((RX_LAST_DESC | RX_ERR_SUMMARY) & ~desc0) == 0) {
579 		dmfep->rx_stats_ierrors += 1;
580 
581 		/*
582 		 * There are some other error bits in the descriptor for
583 		 * which there don't seem to be appropriate MAC statistics,
584 		 * notably RX_COLLISION and perhaps RX_DESC_ERR.  The
585 		 * latter may not be possible if it is supposed to indicate
586 		 * that one buffer has been filled with a partial packet
587 		 * and the next buffer required for the rest of the packet
588 		 * was not available, as all our buffers are more than large
589 		 * enough for a whole packet without fragmenting.
590 		 */
591 
592 		if (desc0 & RX_OVERFLOW) {
593 			dmfep->rx_stats_overflow += 1;
594 
595 		} else if (desc0 & RX_RUNT_FRAME)
596 			dmfep->rx_stats_short += 1;
597 
598 		if (desc0 & RX_CRC)
599 			dmfep->rx_stats_fcs += 1;
600 
601 		if (desc0 & RX_FRAME2LONG)
602 			dmfep->rx_stats_toolong += 1;
603 	}
604 
605 	/*
606 	 * A receive watchdog timeout is counted as a MAC-level receive
607 	 * error.  Strangely, it doesn't set the packet error summary bit,
608 	 * according to the chip data sheet :-?
609 	 */
610 	if (desc0 & RX_RCV_WD_TO)
611 		dmfep->rx_stats_macrcv_errors += 1;
612 
613 	if (desc0 & RX_DRIBBLING)
614 		dmfep->rx_stats_align += 1;
615 
616 	if (desc0 & RX_MII_ERR)
617 		dmfep->rx_stats_macrcv_errors += 1;
618 }
619 
620 /*
621  * Receive incoming packet(s) and pass them up ...
622  */
623 static mblk_t *
624 dmfe_getp(dmfe_t *dmfep)
625 {
626 	dma_area_t *descp;
627 	mblk_t **tail;
628 	mblk_t *head;
629 	mblk_t *mp;
630 	char *rxb;
631 	uchar_t *dp;
632 	uint32_t desc0;
633 	uint32_t misses;
634 	int packet_length;
635 	int index;
636 
637 	mutex_enter(dmfep->rxlock);
638 
639 	/*
640 	 * Update the missed frame statistic from the on-chip counter.
641 	 */
642 	misses = dmfe_chip_get32(dmfep, MISSED_FRAME_REG);
643 	dmfep->rx_stats_norcvbuf += (misses & MISSED_FRAME_MASK);
644 
645 	/*
646 	 * sync (all) receive descriptors before inspecting them
647 	 */
648 	descp = &dmfep->rx_desc;
649 	DMA_SYNC(descp, DDI_DMA_SYNC_FORKERNEL);
650 
651 	/*
652 	 * We should own at least one RX entry, since we've had a
653 	 * receive interrupt, but let's not be dogmatic about it.
654 	 */
655 	index = dmfep->rx.next_free;
656 	desc0 = dmfe_ring_get32(descp, index, DESC0);
657 	if (desc0 & RX_OWN)
658 		DMFE_DEBUG(("dmfe_getp: no work, desc0 0x%x", desc0));
659 
660 	for (head = NULL, tail = &head; (desc0 & RX_OWN) == 0; ) {
661 		/*
662 		 * Maintain statistics for every descriptor returned
663 		 * to us by the chip ...
664 		 */
665 		DMFE_DEBUG(("dmfe_getp: desc0 0x%x", desc0));
666 		dmfe_update_rx_stats(dmfep, desc0);
667 
668 		/*
669 		 * Check that the entry has both "packet start" and
670 		 * "packet end" flags.  We really shouldn't get packet
671 		 * fragments, 'cos all the RX buffers are bigger than
672 		 * the largest valid packet.  So we'll just drop any
673 		 * fragments we find & skip on to the next entry.
674 		 */
675 		if (((RX_FIRST_DESC | RX_LAST_DESC) & ~desc0) != 0) {
676 			DMFE_DEBUG(("dmfe_getp: dropping fragment"));
677 			goto skip;
678 		}
679 
680 		/*
681 		 * A whole packet in one buffer.  We have to check error
682 		 * status and packet length before forwarding it upstream.
683 		 */
684 		if (desc0 & RX_ERR_SUMMARY) {
685 			DMFE_DEBUG(("dmfe_getp: dropping errored packet"));
686 			goto skip;
687 		}
688 
689 		packet_length = (desc0 >> 16) & 0x3fff;
690 		if (packet_length > DMFE_MAX_PKT_SIZE) {
691 			DMFE_DEBUG(("dmfe_getp: dropping oversize packet, "
692 			    "length %d", packet_length));
693 			goto skip;
694 		} else if (packet_length < ETHERMIN) {
695 			/*
696 			 * Note that VLAN packet would be even larger,
697 			 * but we don't worry about dropping runt VLAN
698 			 * frames.
699 			 *
700 			 * This check is probably redundant, as well,
701 			 * since the hardware should drop RUNT frames.
702 			 */
703 			DMFE_DEBUG(("dmfe_getp: dropping undersize packet, "
704 			    "length %d", packet_length));
705 			goto skip;
706 		}
707 
708 		/*
709 		 * Sync the data, so we can examine it; then check that
710 		 * the packet is really intended for us (remember that
711 		 * if we're using Imperfect Filtering, then the chip will
712 		 * receive unicast packets sent to stations whose addresses
713 		 * just happen to hash to the same value as our own; we
714 		 * discard these here so they don't get sent upstream ...)
715 		 */
716 		(void) ddi_dma_sync(dmfep->rx_buff.dma_hdl,
717 		    index * DMFE_BUF_SIZE, DMFE_BUF_SIZE,
718 		    DDI_DMA_SYNC_FORKERNEL);
719 		rxb = &dmfep->rx_buff.mem_va[index*DMFE_BUF_SIZE];
720 
721 
722 		/*
723 		 * We do not bother to check that the packet is really for
724 		 * us, we let the MAC framework make that check instead.
725 		 * This is especially important if we ever want to support
726 		 * multiple MAC addresses.
727 		 */
728 
729 		/*
730 		 * Packet looks good; get a buffer to copy it into.  We
731 		 * allow some space at the front of the allocated buffer
732 		 * (HEADROOM) in case any upstream modules want to prepend
733 		 * some sort of header.  The value has been carefully chosen
734 		 * So that it also has the side-effect of making the packet
735 		 * *contents* 4-byte aligned, as required by NCA!
736 		 */
737 		mp = allocb(DMFE_HEADROOM + packet_length, 0);
738 		if (mp == NULL) {
739 			DMFE_DEBUG(("dmfe_getp: no buffer - dropping packet"));
740 			dmfep->rx_stats_norcvbuf += 1;
741 			goto skip;
742 		}
743 
744 		/*
745 		 * Account for statistics of good packets.
746 		 */
747 		dmfep->rx_stats_ipackets += 1;
748 		dmfep->rx_stats_rbytes += packet_length;
749 		if (desc0 & RX_MULTI_FRAME) {
750 			if (bcmp(rxb, dmfe_broadcast_addr, ETHERADDRL)) {
751 				dmfep->rx_stats_multi += 1;
752 			} else {
753 				dmfep->rx_stats_bcast += 1;
754 			}
755 		}
756 
757 		/*
758 		 * Copy the packet into the STREAMS buffer
759 		 */
760 		dp = mp->b_rptr += DMFE_HEADROOM;
761 		mp->b_cont = mp->b_next = NULL;
762 
763 		/*
764 		 * Don't worry about stripping the vlan tag, the MAC
765 		 * layer will take care of that for us.
766 		 */
767 		bcopy(rxb, dp, packet_length);
768 
769 		/*
770 		 * Fix up the packet length, and link it to the chain
771 		 */
772 		mp->b_wptr = mp->b_rptr + packet_length - ETHERFCSL;
773 		*tail = mp;
774 		tail = &mp->b_next;
775 
776 	skip:
777 		/*
778 		 * Return ownership of ring entry & advance to next
779 		 */
780 		dmfe_ring_put32(descp, index, DESC0, RX_OWN);
781 		index = NEXT(index, dmfep->rx.n_desc);
782 		desc0 = dmfe_ring_get32(descp, index, DESC0);
783 	}
784 
785 	/*
786 	 * Remember where to start looking next time ...
787 	 */
788 	dmfep->rx.next_free = index;
789 
790 	/*
791 	 * sync the receive descriptors that we've given back
792 	 * (actually, we sync all of them for simplicity), and
793 	 * wake the chip in case it had suspended receive
794 	 */
795 	DMA_SYNC(descp, DDI_DMA_SYNC_FORDEV);
796 	dmfe_chip_put32(dmfep, RX_POLL_REG, 0);
797 
798 	mutex_exit(dmfep->rxlock);
799 	return (head);
800 }
801 
802 #undef	DMFE_DBG
803 
804 
805 /*
806  * ========== Primary TX side routines ==========
807  */
808 
809 #define	DMFE_DBG	DMFE_DBG_SEND	/* debug flag for this code	*/
810 
811 /*
812  *	TX ring management:
813  *
814  *	There are <tx.n_desc> entries in the ring, of which those from
815  *	<tx.next_free> round to but not including <tx.next_busy> must
816  *	be owned by the CPU.  The number of such entries should equal
817  *	<tx.n_free>; but there may also be some more entries which the
818  *	chip has given back but which we haven't yet accounted for.
819  *	The routine dmfe_reclaim_tx_desc() adjusts the indexes & counts
820  *	as it discovers such entries.
821  *
822  *	Initially, or when the ring is entirely free:
823  *		C = Owned by CPU
824  *		D = Owned by Davicom (DMFE) chip
825  *
826  *	tx.next_free					tx.n_desc = 16
827  *	  |
828  *	  v
829  *	+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+
830  *	| C | C | C | C | C | C | C | C | C | C | C | C | C | C | C | C |
831  *	+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+
832  *	  ^
833  *	  |
834  *	tx.next_busy					tx.n_free = 16
835  *
836  *	On entry to reclaim() during normal use:
837  *
838  *					tx.next_free	tx.n_desc = 16
839  *					      |
840  *					      v
841  *	+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+
842  *	| C | C | C | C | C | C | D | D | D | C | C | C | C | C | C | C |
843  *	+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+
844  *		  ^
845  *		  |
846  *		tx.next_busy				tx.n_free = 9
847  *
848  *	On exit from reclaim():
849  *
850  *					tx.next_free	tx.n_desc = 16
851  *					      |
852  *					      v
853  *	+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+
854  *	| C | C | C | C | C | C | D | D | D | C | C | C | C | C | C | C |
855  *	+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+
856  *				  ^
857  *				  |
858  *			     tx.next_busy		tx.n_free = 13
859  *
860  *	The ring is considered "full" when only one entry is owned by
861  *	the CPU; thus <tx.n_free> should always be >= 1.
862  *
863  *			tx.next_free			tx.n_desc = 16
864  *			      |
865  *			      v
866  *	+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+
867  *	| D | D | D | D | D | C | D | D | D | D | D | D | D | D | D | D |
868  *	+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+
869  *				  ^
870  *				  |
871  *			     tx.next_busy		tx.n_free = 1
872  */
873 
874 /*
875  * Function to update transmit statistics on various errors
876  */
877 static void
878 dmfe_update_tx_stats(dmfe_t *dmfep, int index, uint32_t desc0, uint32_t desc1)
879 {
880 	uint32_t collisions;
881 	uint32_t errbits;
882 	uint32_t errsum;
883 
884 	ASSERT(mutex_owned(dmfep->txlock));
885 
886 	collisions = ((desc0 >> 3) & 0x0f);
887 	errsum = desc0 & TX_ERR_SUMMARY;
888 	errbits = desc0 & (TX_UNDERFLOW | TX_LATE_COLL | TX_CARRIER_LOSS |
889 	    TX_NO_CARRIER | TX_EXCESS_COLL | TX_JABBER_TO);
890 	if ((errsum == 0) != (errbits == 0)) {
891 		dmfe_log(dmfep, "dubious TX error status 0x%x", desc0);
892 		desc0 |= TX_ERR_SUMMARY;
893 	}
894 
895 	if (desc0 & TX_ERR_SUMMARY) {
896 		dmfep->tx_stats_oerrors += 1;
897 
898 		/*
899 		 * If we ever see a transmit jabber timeout, we count it
900 		 * as a MAC-level transmit error; but we probably won't
901 		 * see it as it causes an Abnormal interrupt and we reset
902 		 * the chip in order to recover
903 		 */
904 		if (desc0 & TX_JABBER_TO) {
905 			dmfep->tx_stats_macxmt_errors += 1;
906 			dmfep->tx_stats_jabber += 1;
907 		}
908 
909 		if (desc0 & TX_UNDERFLOW)
910 			dmfep->tx_stats_underflow += 1;
911 		else if (desc0 & TX_LATE_COLL)
912 			dmfep->tx_stats_xmtlatecoll += 1;
913 
914 		if (desc0 & (TX_CARRIER_LOSS | TX_NO_CARRIER))
915 			dmfep->tx_stats_nocarrier += 1;
916 
917 		if (desc0 & TX_EXCESS_COLL) {
918 			dmfep->tx_stats_excoll += 1;
919 			collisions = 16;
920 		}
921 	} else {
922 		int	bit = index % NBBY;
923 		int	byt = index / NBBY;
924 
925 		if (dmfep->tx_mcast[byt] & bit) {
926 			dmfep->tx_mcast[byt] &= ~bit;
927 			dmfep->tx_stats_multi += 1;
928 
929 		} else if (dmfep->tx_bcast[byt] & bit) {
930 			dmfep->tx_bcast[byt] &= ~bit;
931 			dmfep->tx_stats_bcast += 1;
932 		}
933 
934 		dmfep->tx_stats_opackets += 1;
935 		dmfep->tx_stats_obytes += desc1 & TX_BUFFER_SIZE1;
936 	}
937 
938 	if (collisions == 1)
939 		dmfep->tx_stats_first_coll += 1;
940 	else if (collisions != 0)
941 		dmfep->tx_stats_multi_coll += 1;
942 	dmfep->tx_stats_collisions += collisions;
943 
944 	if (desc0 & TX_DEFERRED)
945 		dmfep->tx_stats_defer += 1;
946 }
947 
948 /*
949  * Reclaim all the ring entries that the chip has returned to us ...
950  *
951  * Returns B_FALSE if no entries could be reclaimed.  Otherwise, reclaims
952  * as many as possible, restarts the TX stall timeout, and returns B_TRUE.
953  */
954 static boolean_t
955 dmfe_reclaim_tx_desc(dmfe_t *dmfep)
956 {
957 	dma_area_t *descp;
958 	uint32_t desc0;
959 	uint32_t desc1;
960 	int i;
961 
962 	ASSERT(mutex_owned(dmfep->txlock));
963 
964 	/*
965 	 * sync transmit descriptor ring before looking at it
966 	 */
967 	descp = &dmfep->tx_desc;
968 	DMA_SYNC(descp, DDI_DMA_SYNC_FORKERNEL);
969 
970 	/*
971 	 * Early exit if there are no descriptors to reclaim, either
972 	 * because they're all reclaimed already, or because the next
973 	 * one is still owned by the chip ...
974 	 */
975 	i = dmfep->tx.next_busy;
976 	if (i == dmfep->tx.next_free)
977 		return (B_FALSE);
978 	desc0 = dmfe_ring_get32(descp, i, DESC0);
979 	if (desc0 & TX_OWN)
980 		return (B_FALSE);
981 
982 	/*
983 	 * Reclaim as many descriptors as possible ...
984 	 */
985 	for (;;) {
986 		desc1 = dmfe_ring_get32(descp, i, DESC1);
987 		ASSERT((desc1 & (TX_SETUP_PACKET | TX_LAST_DESC)) != 0);
988 
989 		if (desc1 & TX_SETUP_PACKET) {
990 			/*
991 			 * Setup packet - restore buffer address
992 			 */
993 			ASSERT(dmfe_ring_get32(descp, i, BUFFER1) ==
994 			    descp->setup_dvma);
995 			dmfe_ring_put32(descp, i, BUFFER1,
996 			    dmfep->tx_buff.mem_dvma + i*DMFE_BUF_SIZE);
997 		} else {
998 			/*
999 			 * Regular packet - just update stats
1000 			 */
1001 			ASSERT(dmfe_ring_get32(descp, i, BUFFER1) ==
1002 			    dmfep->tx_buff.mem_dvma + i*DMFE_BUF_SIZE);
1003 			dmfe_update_tx_stats(dmfep, i, desc0, desc1);
1004 		}
1005 
1006 #if	DMFEDEBUG
1007 		/*
1008 		 * We can use one of the SPARE bits in the TX descriptor
1009 		 * to track when a ring buffer slot is reclaimed.  Then
1010 		 * we can deduce the last operation on a slot from the
1011 		 * top half of DESC0:
1012 		 *
1013 		 *	0x8000 xxxx	given to DMFE chip (TX_OWN)
1014 		 *	0x7fff xxxx	returned but not yet reclaimed
1015 		 *	0x3fff xxxx	reclaimed
1016 		 */
1017 #define	TX_PEND_RECLAIM		(1UL<<30)
1018 		dmfe_ring_put32(descp, i, DESC0, desc0 & ~TX_PEND_RECLAIM);
1019 #endif	/* DMFEDEBUG */
1020 
1021 		/*
1022 		 * Update count & index; we're all done if the ring is
1023 		 * now fully reclaimed, or the next entry if still owned
1024 		 * by the chip ...
1025 		 */
1026 		dmfep->tx.n_free += 1;
1027 		i = NEXT(i, dmfep->tx.n_desc);
1028 		if (i == dmfep->tx.next_free)
1029 			break;
1030 		desc0 = dmfe_ring_get32(descp, i, DESC0);
1031 		if (desc0 & TX_OWN)
1032 			break;
1033 	}
1034 
1035 	dmfep->tx.next_busy = i;
1036 	dmfep->tx_pending_tix = 0;
1037 	return (B_TRUE);
1038 }
1039 
1040 /*
1041  * Send the message in the message block chain <mp>.
1042  *
1043  * The message is freed if and only if its contents are successfully copied
1044  * and queued for transmission (so that the return value is B_TRUE).
1045  * If we can't queue the message, the return value is B_FALSE and
1046  * the message is *not* freed.
1047  *
1048  * This routine handles the special case of <mp> == NULL, which indicates
1049  * that we want to "send" the special "setup packet" allocated during
1050  * startup.  We have to use some different flags in the packet descriptor
1051  * to say its a setup packet (from the global <dmfe_setup_desc1>), and the
1052  * setup packet *isn't* freed after use.
1053  */
1054 static boolean_t
1055 dmfe_send_msg(dmfe_t *dmfep, mblk_t *mp)
1056 {
1057 	dma_area_t *descp;
1058 	mblk_t *bp;
1059 	char *txb;
1060 	uint32_t desc1;
1061 	uint32_t index;
1062 	size_t totlen;
1063 	size_t mblen;
1064 
1065 	/*
1066 	 * If the number of free slots is below the reclaim threshold
1067 	 * (soft limit), we'll try to reclaim some.  If we fail, and
1068 	 * the number of free slots is also below the minimum required
1069 	 * (the hard limit, usually 1), then we can't send the packet.
1070 	 */
1071 	mutex_enter(dmfep->txlock);
1072 	if (dmfep->tx.n_free <= dmfe_tx_reclaim_level &&
1073 	    dmfe_reclaim_tx_desc(dmfep) == B_FALSE &&
1074 	    dmfep->tx.n_free <= dmfe_tx_min_free) {
1075 		/*
1076 		 * Resource shortage - return B_FALSE so the packet
1077 		 * will be queued for retry after the next TX-done
1078 		 * interrupt.
1079 		 */
1080 		mutex_exit(dmfep->txlock);
1081 		DMFE_DEBUG(("dmfe_send_msg: no free descriptors"));
1082 		return (B_FALSE);
1083 	}
1084 
1085 	/*
1086 	 * There's a slot available, so claim it by incrementing
1087 	 * the next-free index and decrementing the free count.
1088 	 * If the ring is currently empty, we also restart the
1089 	 * stall-detect timer.  The ASSERTions check that our
1090 	 * invariants still hold:
1091 	 *	the next-free index must not match the next-busy index
1092 	 *	there must still be at least one free entry
1093 	 * After this, we now have exclusive ownership of the ring
1094 	 * entry (and matching buffer) indicated by <index>, so we
1095 	 * don't need to hold the TX lock any longer
1096 	 */
1097 	index = dmfep->tx.next_free;
1098 	dmfep->tx.next_free = NEXT(index, dmfep->tx.n_desc);
1099 	ASSERT(dmfep->tx.next_free != dmfep->tx.next_busy);
1100 	if (dmfep->tx.n_free-- == dmfep->tx.n_desc)
1101 		dmfep->tx_pending_tix = 0;
1102 	ASSERT(dmfep->tx.n_free >= 1);
1103 	mutex_exit(dmfep->txlock);
1104 
1105 	/*
1106 	 * Check the ownership of the ring entry ...
1107 	 */
1108 	descp = &dmfep->tx_desc;
1109 	ASSERT((dmfe_ring_get32(descp, index, DESC0) & TX_OWN) == 0);
1110 
1111 	if (mp == NULL) {
1112 		/*
1113 		 * Indicates we should send a SETUP packet, which we do by
1114 		 * temporarily switching the BUFFER1 pointer in the ring
1115 		 * entry.  The reclaim routine will restore BUFFER1 to its
1116 		 * usual value.
1117 		 *
1118 		 * Note that as the setup packet is tagged on the end of
1119 		 * the TX ring, when we sync the descriptor we're also
1120 		 * implicitly syncing the setup packet - hence, we don't
1121 		 * need a separate ddi_dma_sync() call here.
1122 		 */
1123 		desc1 = dmfe_setup_desc1;
1124 		dmfe_ring_put32(descp, index, BUFFER1, descp->setup_dvma);
1125 	} else {
1126 		/*
1127 		 * A regular packet; we copy the data into a pre-mapped
1128 		 * buffer, which avoids the overhead (and complication)
1129 		 * of mapping/unmapping STREAMS buffers and keeping hold
1130 		 * of them until the DMA has completed.
1131 		 *
1132 		 * Because all buffers are the same size, and larger
1133 		 * than the longest single valid message, we don't have
1134 		 * to bother about splitting the message across multiple
1135 		 * buffers.
1136 		 */
1137 		txb = &dmfep->tx_buff.mem_va[index*DMFE_BUF_SIZE];
1138 		totlen = 0;
1139 		bp = mp;
1140 
1141 		/*
1142 		 * Copy all (remaining) mblks in the message ...
1143 		 */
1144 		for (; bp != NULL; bp = bp->b_cont) {
1145 			mblen = MBLKL(bp);
1146 			if ((totlen += mblen) <= DMFE_MAX_PKT_SIZE) {
1147 				bcopy(bp->b_rptr, txb, mblen);
1148 				txb += mblen;
1149 			}
1150 		}
1151 
1152 		/*
1153 		 * Is this a multicast or broadcast packet?  We do
1154 		 * this so that we can track statistics accurately
1155 		 * when we reclaim it.
1156 		 */
1157 		txb = &dmfep->tx_buff.mem_va[index*DMFE_BUF_SIZE];
1158 		if (txb[0] & 0x1) {
1159 			if (bcmp(txb, dmfe_broadcast_addr, ETHERADDRL) == 0) {
1160 				dmfep->tx_bcast[index / NBBY] |=
1161 				    (1 << (index % NBBY));
1162 			} else {
1163 				dmfep->tx_mcast[index / NBBY] |=
1164 				    (1 << (index % NBBY));
1165 			}
1166 		}
1167 
1168 		/*
1169 		 * We'e reached the end of the chain; and we should have
1170 		 * collected no more than DMFE_MAX_PKT_SIZE bytes into our
1171 		 * buffer.  Note that the <size> field in the descriptor is
1172 		 * only 11 bits, so bigger packets would be a problem!
1173 		 */
1174 		ASSERT(bp == NULL);
1175 		ASSERT(totlen <= DMFE_MAX_PKT_SIZE);
1176 		totlen &= TX_BUFFER_SIZE1;
1177 		desc1 = TX_FIRST_DESC | TX_LAST_DESC | totlen;
1178 
1179 		(void) ddi_dma_sync(dmfep->tx_buff.dma_hdl,
1180 		    index * DMFE_BUF_SIZE, DMFE_BUF_SIZE, DDI_DMA_SYNC_FORDEV);
1181 	}
1182 
1183 	/*
1184 	 * Update ring descriptor entries, sync them, and wake up the
1185 	 * transmit process
1186 	 */
1187 	if ((index & dmfe_tx_int_factor) == 0)
1188 		desc1 |= TX_INT_ON_COMP;
1189 	desc1 |= TX_CHAINING;
1190 	dmfe_ring_put32(descp, index, DESC1, desc1);
1191 	dmfe_ring_put32(descp, index, DESC0, TX_OWN);
1192 	DMA_SYNC(descp, DDI_DMA_SYNC_FORDEV);
1193 	dmfe_chip_put32(dmfep, TX_POLL_REG, 0);
1194 
1195 	/*
1196 	 * Finally, free the message & return success
1197 	 */
1198 	if (mp)
1199 		freemsg(mp);
1200 	return (B_TRUE);
1201 }
1202 
1203 /*
1204  *	dmfe_m_tx() -- send a chain of packets
1205  *
1206  *	Called when packet(s) are ready to be transmitted. A pointer to an
1207  *	M_DATA message that contains the packet is passed to this routine.
1208  *	The complete LLC header is contained in the message's first message
1209  *	block, and the remainder of the packet is contained within
1210  *	additional M_DATA message blocks linked to the first message block.
1211  *
1212  *	Additional messages may be passed by linking with b_next.
1213  */
1214 static mblk_t *
1215 dmfe_m_tx(void *arg, mblk_t *mp)
1216 {
1217 	dmfe_t *dmfep = arg;			/* private device info	*/
1218 	mblk_t *next;
1219 
1220 	ASSERT(mp != NULL);
1221 	ASSERT(dmfep->mac_state == DMFE_MAC_STARTED);
1222 
1223 	if (dmfep->chip_state != CHIP_RUNNING)
1224 		return (mp);
1225 
1226 	while (mp != NULL) {
1227 		next = mp->b_next;
1228 		mp->b_next = NULL;
1229 		if (!dmfe_send_msg(dmfep, mp)) {
1230 			mp->b_next = next;
1231 			break;
1232 		}
1233 		mp = next;
1234 	}
1235 
1236 	return (mp);
1237 }
1238 
1239 #undef	DMFE_DBG
1240 
1241 
1242 /*
1243  * ========== Address-setting routines (TX-side) ==========
1244  */
1245 
1246 #define	DMFE_DBG	DMFE_DBG_ADDR	/* debug flag for this code	*/
1247 
1248 /*
1249  * Find the index of the relevant bit in the setup packet.
1250  * This must mirror the way the hardware will actually calculate it!
1251  */
1252 static uint32_t
1253 dmfe_hash_index(const uint8_t *address)
1254 {
1255 	uint32_t const POLY = HASH_POLY;
1256 	uint32_t crc = HASH_CRC;
1257 	uint32_t index;
1258 	uint32_t msb;
1259 	uchar_t currentbyte;
1260 	int byteslength;
1261 	int shift;
1262 	int bit;
1263 
1264 	for (byteslength = 0; byteslength < ETHERADDRL; ++byteslength) {
1265 		currentbyte = address[byteslength];
1266 		for (bit = 0; bit < 8; ++bit) {
1267 			msb = crc >> 31;
1268 			crc <<= 1;
1269 			if (msb ^ (currentbyte & 1)) {
1270 				crc ^= POLY;
1271 				crc |= 0x00000001;
1272 			}
1273 			currentbyte >>= 1;
1274 		}
1275 	}
1276 
1277 	for (index = 0, bit = 23, shift = 8; shift >= 0; ++bit, --shift)
1278 		index |= (((crc >> bit) & 1) << shift);
1279 
1280 	return (index);
1281 }
1282 
1283 /*
1284  * Find and set/clear the relevant bit in the setup packet hash table
1285  * This must mirror the way the hardware will actually interpret it!
1286  */
1287 static void
1288 dmfe_update_hash(dmfe_t *dmfep, uint32_t index, boolean_t val)
1289 {
1290 	dma_area_t *descp;
1291 	uint32_t tmp;
1292 
1293 	ASSERT(mutex_owned(dmfep->oplock));
1294 
1295 	descp = &dmfep->tx_desc;
1296 	tmp = dmfe_setup_get32(descp, index/16);
1297 	if (val)
1298 		tmp |= 1 << (index%16);
1299 	else
1300 		tmp &= ~(1 << (index%16));
1301 	dmfe_setup_put32(descp, index/16, tmp);
1302 }
1303 
1304 /*
1305  * Update the refcount for the bit in the setup packet corresponding
1306  * to the specified address; if it changes between zero & nonzero,
1307  * also update the bitmap itself & return B_TRUE, so that the caller
1308  * knows to re-send the setup packet.  Otherwise (only the refcount
1309  * changed), return B_FALSE
1310  */
1311 static boolean_t
1312 dmfe_update_mcast(dmfe_t *dmfep, const uint8_t *mca, boolean_t val)
1313 {
1314 	uint32_t index;
1315 	uint8_t *refp;
1316 	boolean_t change;
1317 
1318 	index = dmfe_hash_index(mca);
1319 	refp = &dmfep->mcast_refs[index];
1320 	change = (val ? (*refp)++ : --(*refp)) == 0;
1321 
1322 	if (change)
1323 		dmfe_update_hash(dmfep, index, val);
1324 
1325 	return (change);
1326 }
1327 
1328 /*
1329  * "Transmit" the (possibly updated) magic setup packet
1330  */
1331 static int
1332 dmfe_send_setup(dmfe_t *dmfep)
1333 {
1334 	int status;
1335 
1336 	ASSERT(mutex_owned(dmfep->oplock));
1337 
1338 	/*
1339 	 * If the chip isn't running, we can't really send the setup frame
1340 	 * now but it doesn't matter, 'cos it will be sent when the transmit
1341 	 * process is restarted (see dmfe_start()).
1342 	 */
1343 	if ((dmfep->opmode & START_TRANSMIT) == 0)
1344 		return (0);
1345 
1346 	/*
1347 	 * "Send" the setup frame.  If it fails (e.g. no resources),
1348 	 * set a flag; then the factotum will retry the "send".  Once
1349 	 * it works, we can clear the flag no matter how many attempts
1350 	 * had previously failed.  We tell the caller that it worked
1351 	 * whether it did or not; after all, it *will* work eventually.
1352 	 */
1353 	status = dmfe_send_msg(dmfep, NULL);
1354 	dmfep->need_setup = status ? B_FALSE : B_TRUE;
1355 	return (0);
1356 }
1357 
1358 /*
1359  *	dmfe_m_unicst() -- set the physical network address
1360  */
1361 static int
1362 dmfe_m_unicst(void *arg, const uint8_t *macaddr)
1363 {
1364 	dmfe_t *dmfep = arg;
1365 	int status;
1366 	int index;
1367 
1368 	/*
1369 	 * Update our current address and send out a new setup packet
1370 	 *
1371 	 * Here we accommodate the use of HASH_ONLY or HASH_AND_PERFECT
1372 	 * filtering modes (we don't support PERFECT_ONLY or INVERSE modes).
1373 	 *
1374 	 * It is said that there is a bug in the 21140 where it fails to
1375 	 * receive packes addresses to the specified perfect filter address.
1376 	 * If the same bug is present in the DM9102A, the TX_FILTER_TYPE1
1377 	 * bit should be set in the module variable dmfe_setup_desc1.
1378 	 *
1379 	 * If TX_FILTER_TYPE1 is set, we will use HASH_ONLY filtering.
1380 	 * In this mode, *all* incoming addresses are hashed and looked
1381 	 * up in the bitmap described by the setup packet.  Therefore,
1382 	 * the bit representing the station address has to be added to
1383 	 * the table before sending it out.  If the address is changed,
1384 	 * the old entry should be removed before the new entry is made.
1385 	 *
1386 	 * NOTE: in this mode, unicast packets that are not intended for
1387 	 * this station may be received; it is up to software to filter
1388 	 * them out afterwards!
1389 	 *
1390 	 * If TX_FILTER_TYPE1 is *not* set, we will use HASH_AND_PERFECT
1391 	 * filtering.  In this mode, multicast addresses are hashed and
1392 	 * checked against the bitmap, while unicast addresses are simply
1393 	 * matched against the one physical address specified in the setup
1394 	 * packet.  This means that we shouldn't receive unicast packets
1395 	 * that aren't intended for us (but software still has to filter
1396 	 * multicast packets just the same).
1397 	 *
1398 	 * Whichever mode we're using, we have to enter the broadcast
1399 	 * address into the multicast filter map too, so we do this on
1400 	 * the first time through after attach or reset.
1401 	 */
1402 	mutex_enter(dmfep->oplock);
1403 
1404 	if (dmfep->addr_set && dmfe_setup_desc1 & TX_FILTER_TYPE1)
1405 		(void) dmfe_update_mcast(dmfep, dmfep->curr_addr, B_FALSE);
1406 	if (dmfe_setup_desc1 & TX_FILTER_TYPE1)
1407 		(void) dmfe_update_mcast(dmfep, macaddr, B_TRUE);
1408 	if (!dmfep->addr_set)
1409 		(void) dmfe_update_mcast(dmfep, dmfe_broadcast_addr, B_TRUE);
1410 
1411 	/*
1412 	 * Remember the new current address
1413 	 */
1414 	ethaddr_copy(macaddr, dmfep->curr_addr);
1415 	dmfep->addr_set = B_TRUE;
1416 
1417 	/*
1418 	 * Install the new physical address into the proper position in
1419 	 * the setup frame; this is only used if we select hash+perfect
1420 	 * filtering, but we'll put it in anyway.  The ugliness here is
1421 	 * down to the usual war of the egg :(
1422 	 */
1423 	for (index = 0; index < ETHERADDRL; index += 2)
1424 		dmfe_setup_put32(&dmfep->tx_desc, SETUPBUF_PHYS+index/2,
1425 		    (macaddr[index+1] << 8) | macaddr[index]);
1426 
1427 	/*
1428 	 * Finally, we're ready to "transmit" the setup frame
1429 	 */
1430 	status = dmfe_send_setup(dmfep);
1431 	mutex_exit(dmfep->oplock);
1432 
1433 	return (status);
1434 }
1435 
1436 /*
1437  *	dmfe_m_multicst() -- enable or disable a multicast address
1438  *
1439  *	Program the hardware to enable/disable the multicast address
1440  *	in "mca" (enable if add is true, otherwise disable it.)
1441  *	We keep a refcount for each bit in the map, so that it still
1442  *	works out properly if multiple addresses hash to the same bit.
1443  *	dmfe_update_mcast() tells us whether the map actually changed;
1444  *	if so, we have to re-"transmit" the magic setup packet.
1445  */
1446 static int
1447 dmfe_m_multicst(void *arg, boolean_t add, const uint8_t *mca)
1448 {
1449 	dmfe_t *dmfep = arg;			/* private device info	*/
1450 	int status = 0;
1451 
1452 	mutex_enter(dmfep->oplock);
1453 	if (dmfe_update_mcast(dmfep, mca, add))
1454 		status = dmfe_send_setup(dmfep);
1455 	mutex_exit(dmfep->oplock);
1456 
1457 	return (status);
1458 }
1459 
1460 #undef	DMFE_DBG
1461 
1462 
1463 /*
1464  * ========== Internal state management entry points ==========
1465  */
1466 
1467 #define	DMFE_DBG	DMFE_DBG_GLD	/* debug flag for this code	*/
1468 
1469 /*
1470  * These routines provide all the functionality required by the
1471  * corresponding MAC layer entry points, but don't update the MAC layer state
1472  * so they can be called internally without disturbing our record
1473  * of what MAC layer thinks we should be doing ...
1474  */
1475 
1476 /*
1477  *	dmfe_stop() -- stop processing, don't reset h/w or rings
1478  */
1479 static void
1480 dmfe_stop(dmfe_t *dmfep)
1481 {
1482 	ASSERT(mutex_owned(dmfep->oplock));
1483 
1484 	dmfe_stop_chip(dmfep, CHIP_STOPPED);
1485 }
1486 
1487 /*
1488  *	dmfe_reset() -- stop processing, reset h/w & rings to initial state
1489  */
1490 static void
1491 dmfe_reset(dmfe_t *dmfep)
1492 {
1493 	ASSERT(mutex_owned(dmfep->oplock));
1494 	ASSERT(mutex_owned(dmfep->rxlock));
1495 	ASSERT(mutex_owned(dmfep->txlock));
1496 
1497 	dmfe_stop_chip(dmfep, CHIP_RESET);
1498 	dmfe_init_rings(dmfep);
1499 }
1500 
1501 /*
1502  *	dmfe_start() -- start transmitting/receiving
1503  */
1504 static void
1505 dmfe_start(dmfe_t *dmfep)
1506 {
1507 	uint32_t gpsr;
1508 
1509 	ASSERT(mutex_owned(dmfep->oplock));
1510 
1511 	ASSERT(dmfep->chip_state == CHIP_RESET ||
1512 	    dmfep->chip_state == CHIP_STOPPED);
1513 
1514 	/*
1515 	 * Make opmode consistent with PHY duplex setting
1516 	 */
1517 	gpsr = dmfe_chip_get32(dmfep, PHY_STATUS_REG);
1518 	if (gpsr & GPS_FULL_DUPLEX)
1519 		dmfep->opmode |= FULL_DUPLEX;
1520 	else
1521 		dmfep->opmode &= ~FULL_DUPLEX;
1522 
1523 	/*
1524 	 * Start transmit processing
1525 	 * Set up the address filters
1526 	 * Start receive processing
1527 	 * Enable interrupts
1528 	 */
1529 	dmfe_start_chip(dmfep, START_TRANSMIT);
1530 	(void) dmfe_send_setup(dmfep);
1531 	drv_usecwait(10);
1532 	dmfe_start_chip(dmfep, START_RECEIVE);
1533 	dmfe_enable_interrupts(dmfep);
1534 }
1535 
1536 /*
1537  * dmfe_restart - restart transmitting/receiving after error or suspend
1538  */
1539 static void
1540 dmfe_restart(dmfe_t *dmfep)
1541 {
1542 	ASSERT(mutex_owned(dmfep->oplock));
1543 
1544 	/*
1545 	 * You need not only <oplock>, but also <rxlock> AND <txlock>
1546 	 * in order to reset the rings, but then <txlock> *mustn't*
1547 	 * be held across the call to dmfe_start()
1548 	 */
1549 	mutex_enter(dmfep->rxlock);
1550 	mutex_enter(dmfep->txlock);
1551 	dmfe_reset(dmfep);
1552 	mutex_exit(dmfep->txlock);
1553 	mutex_exit(dmfep->rxlock);
1554 	if (dmfep->mac_state == DMFE_MAC_STARTED)
1555 		dmfe_start(dmfep);
1556 }
1557 
1558 
1559 /*
1560  * ========== MAC-required management entry points ==========
1561  */
1562 
1563 /*
1564  *	dmfe_m_stop() -- stop transmitting/receiving
1565  */
1566 static void
1567 dmfe_m_stop(void *arg)
1568 {
1569 	dmfe_t *dmfep = arg;			/* private device info	*/
1570 
1571 	/*
1572 	 * Just stop processing, then record new MAC state
1573 	 */
1574 	mutex_enter(dmfep->oplock);
1575 	dmfe_stop(dmfep);
1576 	dmfep->mac_state = DMFE_MAC_STOPPED;
1577 	mutex_exit(dmfep->oplock);
1578 }
1579 
1580 /*
1581  *	dmfe_m_start() -- start transmitting/receiving
1582  */
1583 static int
1584 dmfe_m_start(void *arg)
1585 {
1586 	dmfe_t *dmfep = arg;			/* private device info	*/
1587 
1588 	/*
1589 	 * Start processing and record new MAC state
1590 	 */
1591 	mutex_enter(dmfep->oplock);
1592 	dmfe_start(dmfep);
1593 	dmfep->mac_state = DMFE_MAC_STARTED;
1594 	mutex_exit(dmfep->oplock);
1595 
1596 	return (0);
1597 }
1598 
1599 /*
1600  * dmfe_m_promisc() -- set or reset promiscuous mode on the board
1601  *
1602  *	Program the hardware to enable/disable promiscuous and/or
1603  *	receive-all-multicast modes.  Davicom don't document this
1604  *	clearly, but it looks like we can do this on-the-fly (i.e.
1605  *	without stopping & restarting the TX/RX processes).
1606  */
1607 static int
1608 dmfe_m_promisc(void *arg, boolean_t on)
1609 {
1610 	dmfe_t *dmfep = arg;
1611 
1612 	mutex_enter(dmfep->oplock);
1613 	dmfep->opmode &= ~(PROMISC_MODE | PASS_MULTICAST);
1614 	if (on)
1615 		dmfep->opmode |= PROMISC_MODE;
1616 	dmfe_set_opmode(dmfep);
1617 	mutex_exit(dmfep->oplock);
1618 
1619 	return (0);
1620 }
1621 
1622 #undef	DMFE_DBG
1623 
1624 
1625 /*
1626  * ========== Factotum, implemented as a softint handler ==========
1627  */
1628 
1629 #define	DMFE_DBG	DMFE_DBG_FACT	/* debug flag for this code	*/
1630 
1631 /*
1632  * The factotum is woken up when there's something to do that we'd rather
1633  * not do from inside a (high-level?) hardware interrupt handler.  Its
1634  * two main tasks are:
1635  *	reset & restart the chip after an error
1636  *	update & restart the chip after a link status change
1637  */
1638 static uint_t
1639 dmfe_factotum(caddr_t arg)
1640 {
1641 	dmfe_t *dmfep;
1642 
1643 	dmfep = (void *)arg;
1644 	ASSERT(dmfep->dmfe_guard == DMFE_GUARD);
1645 
1646 	mutex_enter(dmfep->oplock);
1647 
1648 	dmfep->factotum_flag = 0;
1649 	DRV_KS_INC(dmfep, KS_FACTOTUM_RUN);
1650 
1651 	/*
1652 	 * Check for chip error ...
1653 	 */
1654 	if (dmfep->chip_state == CHIP_ERROR) {
1655 		/*
1656 		 * Error recovery required: reset the chip and the rings,
1657 		 * then, if it's supposed to be running, kick it off again.
1658 		 */
1659 		DRV_KS_INC(dmfep, KS_RECOVERY);
1660 		dmfe_restart(dmfep);
1661 	} else if (dmfep->need_setup) {
1662 		(void) dmfe_send_setup(dmfep);
1663 	}
1664 	mutex_exit(dmfep->oplock);
1665 
1666 	/*
1667 	 * Then, check the link state.  We need <milock> but not <oplock>
1668 	 * to do this, but if something's changed, we need <oplock> as well
1669 	 * in order to stop/restart the chip!  Note: we could simply hold
1670 	 * <oplock> right through here, but we'd rather not 'cos checking
1671 	 * the link state involves reading over the bit-serial MII bus,
1672 	 * which takes ~500us even when nothing's changed.  Holding <oplock>
1673 	 * would lock out the interrupt handler for the duration, so it's
1674 	 * better to release it first and reacquire it only if needed.
1675 	 */
1676 	mutex_enter(dmfep->milock);
1677 	if (dmfe_check_link(dmfep)) {
1678 		mutex_enter(dmfep->oplock);
1679 		dmfe_stop(dmfep);
1680 		DRV_KS_INC(dmfep, KS_LINK_CHECK);
1681 		if (dmfep->update_phy) {
1682 			/*
1683 			 *  The chip may reset itself for some unknown
1684 			 * reason.  If this happens, the chip will use
1685 			 * default settings (for speed, duplex, and autoneg),
1686 			 * which possibly aren't the user's desired settings.
1687 			 */
1688 			dmfe_update_phy(dmfep);
1689 			dmfep->update_phy = B_FALSE;
1690 		}
1691 		dmfe_recheck_link(dmfep, B_FALSE);
1692 		if (dmfep->mac_state == DMFE_MAC_STARTED)
1693 			dmfe_start(dmfep);
1694 		mutex_exit(dmfep->oplock);
1695 	}
1696 	mutex_exit(dmfep->milock);
1697 
1698 	/*
1699 	 * Keep MAC up-to-date about the state of the link ...
1700 	 */
1701 	mac_link_update(dmfep->mh, dmfep->link_state);
1702 
1703 	return (DDI_INTR_CLAIMED);
1704 }
1705 
1706 static void
1707 dmfe_wake_factotum(dmfe_t *dmfep, int ks_id, const char *why)
1708 {
1709 	DMFE_DEBUG(("dmfe_wake_factotum: %s [%d] flag %d",
1710 	    why, ks_id, dmfep->factotum_flag));
1711 
1712 	ASSERT(mutex_owned(dmfep->oplock));
1713 	DRV_KS_INC(dmfep, ks_id);
1714 
1715 	if (dmfep->factotum_flag++ == 0)
1716 		ddi_trigger_softintr(dmfep->factotum_id);
1717 }
1718 
1719 #undef	DMFE_DBG
1720 
1721 
1722 /*
1723  * ========== Periodic Tasks (Cyclic handler & friends) ==========
1724  */
1725 
1726 #define	DMFE_DBG	DMFE_DBG_TICK	/* debug flag for this code	*/
1727 
1728 /*
1729  * Periodic tick tasks, run from the cyclic handler
1730  *
1731  * Check the state of the link and wake the factotum if necessary
1732  */
1733 static void
1734 dmfe_tick_link_check(dmfe_t *dmfep, uint32_t gpsr, uint32_t istat)
1735 {
1736 	link_state_t phy_state;
1737 	link_state_t utp_state;
1738 	const char *why;
1739 	int ks_id;
1740 
1741 	_NOTE(ARGUNUSED(istat))
1742 
1743 	ASSERT(mutex_owned(dmfep->oplock));
1744 
1745 	/*
1746 	 * Is it time to wake the factotum?  We do so periodically, in
1747 	 * case the fast check below doesn't always reveal a link change
1748 	 */
1749 	if (dmfep->link_poll_tix-- == 0) {
1750 		dmfep->link_poll_tix = factotum_tix;
1751 		why = "tick (link poll)";
1752 		ks_id = KS_TICK_LINK_POLL;
1753 	} else {
1754 		why = NULL;
1755 		ks_id = KS_TICK_LINK_STATE;
1756 	}
1757 
1758 	/*
1759 	 * Has the link status changed?  If so, we might want to wake
1760 	 * the factotum to deal with it.
1761 	 */
1762 	phy_state = (gpsr & GPS_LINK_STATUS) ? LINK_STATE_UP : LINK_STATE_DOWN;
1763 	utp_state = (gpsr & GPS_UTP_SIG) ? LINK_STATE_UP : LINK_STATE_DOWN;
1764 	if (phy_state != utp_state)
1765 		why = "tick (phy <> utp)";
1766 	else if ((dmfep->link_state == LINK_STATE_UP) &&
1767 	    (phy_state == LINK_STATE_DOWN))
1768 		why = "tick (UP -> DOWN)";
1769 	else if (phy_state != dmfep->link_state) {
1770 		if (dmfep->link_poll_tix > factotum_fast_tix)
1771 			dmfep->link_poll_tix = factotum_fast_tix;
1772 	}
1773 
1774 	if (why != NULL) {
1775 		DMFE_DEBUG(("dmfe_%s: link %d phy %d utp %d",
1776 		    why, dmfep->link_state, phy_state, utp_state));
1777 		dmfe_wake_factotum(dmfep, ks_id, why);
1778 	}
1779 }
1780 
1781 /*
1782  * Periodic tick tasks, run from the cyclic handler
1783  *
1784  * Check for TX stall; flag an error and wake the factotum if so.
1785  */
1786 static void
1787 dmfe_tick_stall_check(dmfe_t *dmfep, uint32_t gpsr, uint32_t istat)
1788 {
1789 	boolean_t tx_stall;
1790 	uint32_t tx_state;
1791 	uint32_t limit;
1792 
1793 	ASSERT(mutex_owned(dmfep->oplock));
1794 
1795 	/*
1796 	 * Check for transmit stall ...
1797 	 *
1798 	 * IF there's at least one packet in the ring, AND the timeout
1799 	 * has elapsed, AND we can't reclaim any descriptors, THEN we've
1800 	 * stalled; we return B_TRUE to trigger a reset-and-recover cycle.
1801 	 *
1802 	 * Note that the timeout limit is based on the transmit engine
1803 	 * state; we allow the transmitter longer to make progress in
1804 	 * some states than in others, based on observations of this
1805 	 * chip's actual behaviour in the lab.
1806 	 *
1807 	 * By observation, we find that on about 1 in 10000 passes through
1808 	 * here, the TX lock is already held.  In that case, we'll skip
1809 	 * the check on this pass rather than wait.  Most likely, the send
1810 	 * routine was holding the lock when the interrupt happened, and
1811 	 * we'll succeed next time through.  In the event of a real stall,
1812 	 * the TX ring will fill up, after which the send routine won't be
1813 	 * called any more and then we're sure to get in.
1814 	 */
1815 	tx_stall = B_FALSE;
1816 	if (mutex_tryenter(dmfep->txlock)) {
1817 		if (dmfep->tx.n_free < dmfep->tx.n_desc) {
1818 			tx_state = TX_PROCESS_STATE(istat);
1819 			if (gpsr & GPS_LINK_100)
1820 				limit = stall_100_tix[tx_state];
1821 			else
1822 				limit = stall_10_tix[tx_state];
1823 			if (++dmfep->tx_pending_tix >= limit &&
1824 			    dmfe_reclaim_tx_desc(dmfep) == B_FALSE) {
1825 				dmfe_log(dmfep, "TX stall detected "
1826 				    "after %d ticks in state %d; "
1827 				    "automatic recovery initiated",
1828 				    dmfep->tx_pending_tix, tx_state);
1829 				tx_stall = B_TRUE;
1830 			}
1831 		}
1832 		mutex_exit(dmfep->txlock);
1833 	}
1834 
1835 	if (tx_stall) {
1836 		dmfe_stop_chip(dmfep, CHIP_ERROR);
1837 		dmfe_wake_factotum(dmfep, KS_TX_STALL, "tick (TX stall)");
1838 	}
1839 }
1840 
1841 /*
1842  * Cyclic callback handler
1843  */
1844 static void
1845 dmfe_cyclic(void *arg)
1846 {
1847 	dmfe_t *dmfep = arg;			/* private device info */
1848 	uint32_t istat;
1849 	uint32_t gpsr;
1850 
1851 	/*
1852 	 * If the chip's not RUNNING, there's nothing to do.
1853 	 * If we can't get the mutex straight away, we'll just
1854 	 * skip this pass; we'll back back soon enough anyway.
1855 	 */
1856 	if (dmfep->chip_state != CHIP_RUNNING)
1857 		return;
1858 	if (mutex_tryenter(dmfep->oplock) == 0)
1859 		return;
1860 
1861 	/*
1862 	 * Recheck chip state (it might have been stopped since we
1863 	 * checked above).  If still running, call each of the *tick*
1864 	 * tasks.  They will check for link change, TX stall, etc ...
1865 	 */
1866 	if (dmfep->chip_state == CHIP_RUNNING) {
1867 		istat = dmfe_chip_get32(dmfep, STATUS_REG);
1868 		gpsr = dmfe_chip_get32(dmfep, PHY_STATUS_REG);
1869 		dmfe_tick_link_check(dmfep, gpsr, istat);
1870 		dmfe_tick_stall_check(dmfep, gpsr, istat);
1871 	}
1872 
1873 	DRV_KS_INC(dmfep, KS_CYCLIC_RUN);
1874 	mutex_exit(dmfep->oplock);
1875 }
1876 
1877 #undef	DMFE_DBG
1878 
1879 
1880 /*
1881  * ========== Hardware interrupt handler ==========
1882  */
1883 
1884 #define	DMFE_DBG	DMFE_DBG_INT	/* debug flag for this code	*/
1885 
1886 /*
1887  *	dmfe_interrupt() -- handle chip interrupts
1888  */
1889 static uint_t
1890 dmfe_interrupt(caddr_t arg)
1891 {
1892 	dmfe_t *dmfep;			/* private device info */
1893 	uint32_t interrupts;
1894 	uint32_t istat;
1895 	const char *msg;
1896 	mblk_t *mp;
1897 	boolean_t warning_msg = B_TRUE;
1898 
1899 	dmfep = (void *)arg;
1900 
1901 	/*
1902 	 * A quick check as to whether the interrupt was from this
1903 	 * device, before we even finish setting up all our local
1904 	 * variables.  Note that reading the interrupt status register
1905 	 * doesn't have any unpleasant side effects such as clearing
1906 	 * the bits read, so it's quite OK to re-read it once we have
1907 	 * determined that we are going to service this interrupt and
1908 	 * grabbed the mutexen.
1909 	 */
1910 	istat = dmfe_chip_get32(dmfep, STATUS_REG);
1911 	if ((istat & (NORMAL_SUMMARY_INT | ABNORMAL_SUMMARY_INT)) == 0)
1912 		return (DDI_INTR_UNCLAIMED);
1913 
1914 	/*
1915 	 * Unfortunately, there can be a race condition between attach()
1916 	 * adding the interrupt handler and initialising the mutexen,
1917 	 * and the handler itself being called because of a pending
1918 	 * interrupt.  So, we check <imask>; if it shows that interrupts
1919 	 * haven't yet been enabled (and therefore we shouldn't really
1920 	 * be here at all), we will just write back the value read from
1921 	 * the status register, thus acknowledging (and clearing) *all*
1922 	 * pending conditions without really servicing them, and claim
1923 	 * the interrupt.
1924 	 */
1925 	if (dmfep->imask == 0) {
1926 		DMFE_DEBUG(("dmfe_interrupt: early interrupt 0x%x", istat));
1927 		dmfe_chip_put32(dmfep, STATUS_REG, istat);
1928 		return (DDI_INTR_CLAIMED);
1929 	}
1930 
1931 	/*
1932 	 * We're committed to servicing this interrupt, but we
1933 	 * need to get the lock before going any further ...
1934 	 */
1935 	mutex_enter(dmfep->oplock);
1936 	DRV_KS_INC(dmfep, KS_INTERRUPT);
1937 
1938 	/*
1939 	 * Identify bits that represent enabled interrupts ...
1940 	 */
1941 	istat |= dmfe_chip_get32(dmfep, STATUS_REG);
1942 	interrupts = istat & dmfep->imask;
1943 	ASSERT(interrupts != 0);
1944 
1945 	DMFE_DEBUG(("dmfe_interrupt: istat 0x%x -> 0x%x", istat, interrupts));
1946 
1947 	/*
1948 	 * Check for any interrupts other than TX/RX done.
1949 	 * If there are any, they are considered Abnormal
1950 	 * and will cause the chip to be reset.
1951 	 */
1952 	if (interrupts & ~(RX_PKTDONE_INT | TX_PKTDONE_INT)) {
1953 		if (istat & ABNORMAL_SUMMARY_INT) {
1954 			/*
1955 			 * Any Abnormal interrupts will lead to us
1956 			 * resetting the chip, so we don't bother
1957 			 * to clear each interrupt individually.
1958 			 *
1959 			 * Our main task here is to identify the problem,
1960 			 * by pointing out the most significant unexpected
1961 			 * bit.  Additional bits may well be consequences
1962 			 * of the first problem, so we consider the possible
1963 			 * causes in order of severity.
1964 			 */
1965 			if (interrupts & SYSTEM_ERR_INT) {
1966 				switch (istat & SYSTEM_ERR_BITS) {
1967 				case SYSTEM_ERR_M_ABORT:
1968 					msg = "Bus Master Abort";
1969 					break;
1970 
1971 				case SYSTEM_ERR_T_ABORT:
1972 					msg = "Bus Target Abort";
1973 					break;
1974 
1975 				case SYSTEM_ERR_PARITY:
1976 					msg = "Parity Error";
1977 					break;
1978 
1979 				default:
1980 					msg = "Unknown System Bus Error";
1981 					break;
1982 				}
1983 			} else if (interrupts & RX_STOPPED_INT) {
1984 				msg = "RX process stopped";
1985 			} else if (interrupts & RX_UNAVAIL_INT) {
1986 				msg = "RX buffer unavailable";
1987 				warning_msg = B_FALSE;
1988 			} else if (interrupts & RX_WATCHDOG_INT) {
1989 				msg = "RX watchdog timeout?";
1990 			} else if (interrupts & RX_EARLY_INT) {
1991 				msg = "RX early interrupt?";
1992 			} else if (interrupts & TX_STOPPED_INT) {
1993 				msg = "TX process stopped";
1994 			} else if (interrupts & TX_JABBER_INT) {
1995 				msg = "TX jabber timeout";
1996 			} else if (interrupts & TX_UNDERFLOW_INT) {
1997 				msg = "TX underflow?";
1998 			} else if (interrupts & TX_EARLY_INT) {
1999 				msg = "TX early interrupt?";
2000 
2001 			} else if (interrupts & LINK_STATUS_INT) {
2002 				msg = "Link status change?";
2003 			} else if (interrupts & GP_TIMER_INT) {
2004 				msg = "Timer expired?";
2005 			}
2006 
2007 			if (warning_msg)
2008 				dmfe_warning(dmfep, "abnormal interrupt, "
2009 				    "status 0x%x: %s", istat, msg);
2010 
2011 			/*
2012 			 * We don't want to run the entire reinitialisation
2013 			 * code out of this (high-level?) interrupt, so we
2014 			 * simply STOP the chip, and wake up the factotum
2015 			 * to reinitalise it ...
2016 			 */
2017 			dmfe_stop_chip(dmfep, CHIP_ERROR);
2018 			dmfe_wake_factotum(dmfep, KS_CHIP_ERROR,
2019 			    "interrupt (error)");
2020 		} else {
2021 			/*
2022 			 * We shouldn't really get here (it would mean
2023 			 * there were some unprocessed enabled bits but
2024 			 * they weren't Abnormal?), but we'll check just
2025 			 * in case ...
2026 			 */
2027 			DMFE_DEBUG(("unexpected interrupt bits: 0x%x", istat));
2028 		}
2029 	}
2030 
2031 	/*
2032 	 * Acknowledge all the original bits - except in the case of an
2033 	 * error, when we leave them unacknowledged so that the recovery
2034 	 * code can see what was going on when the problem occurred ...
2035 	 */
2036 	if (dmfep->chip_state != CHIP_ERROR) {
2037 		(void) dmfe_chip_put32(dmfep, STATUS_REG, istat);
2038 		/*
2039 		 * Read-after-write forces completion on PCI bus.
2040 		 *
2041 		 */
2042 		(void) dmfe_chip_get32(dmfep, STATUS_REG);
2043 	}
2044 
2045 
2046 	/*
2047 	 * We've finished talking to the chip, so we can drop <oplock>
2048 	 * before handling the normal interrupts, which only involve
2049 	 * manipulation of descriptors ...
2050 	 */
2051 	mutex_exit(dmfep->oplock);
2052 
2053 	if (interrupts & RX_PKTDONE_INT)
2054 		if ((mp = dmfe_getp(dmfep)) != NULL)
2055 			mac_rx(dmfep->mh, NULL, mp);
2056 
2057 	if (interrupts & TX_PKTDONE_INT) {
2058 		/*
2059 		 * The only reason for taking this interrupt is to give
2060 		 * MAC a chance to schedule queued packets after a
2061 		 * ring-full condition.  To minimise the number of
2062 		 * redundant TX-Done interrupts, we only mark two of the
2063 		 * ring descriptors as 'interrupt-on-complete' - all the
2064 		 * others are simply handed back without an interrupt.
2065 		 */
2066 		if (dmfe_reclaim_on_done && mutex_tryenter(dmfep->txlock)) {
2067 			(void) dmfe_reclaim_tx_desc(dmfep);
2068 			mutex_exit(dmfep->txlock);
2069 		}
2070 		mac_tx_update(dmfep->mh);
2071 	}
2072 
2073 	return (DDI_INTR_CLAIMED);
2074 }
2075 
2076 #undef	DMFE_DBG
2077 
2078 
2079 /*
2080  * ========== Statistics update handler ==========
2081  */
2082 
2083 #define	DMFE_DBG	DMFE_DBG_STATS	/* debug flag for this code	*/
2084 
2085 static int
2086 dmfe_m_stat(void *arg, uint_t stat, uint64_t *val)
2087 {
2088 	dmfe_t *dmfep = arg;
2089 	int rv = 0;
2090 
2091 	mutex_enter(dmfep->milock);
2092 	mutex_enter(dmfep->oplock);
2093 	mutex_enter(dmfep->rxlock);
2094 	mutex_enter(dmfep->txlock);
2095 
2096 	/* make sure we have all the stats collected */
2097 	(void) dmfe_reclaim_tx_desc(dmfep);
2098 
2099 	switch (stat) {
2100 	case MAC_STAT_IFSPEED:
2101 		*val = dmfep->op_stats_speed;
2102 		break;
2103 
2104 	case MAC_STAT_IPACKETS:
2105 		*val = dmfep->rx_stats_ipackets;
2106 		break;
2107 
2108 	case MAC_STAT_MULTIRCV:
2109 		*val = dmfep->rx_stats_multi;
2110 		break;
2111 
2112 	case MAC_STAT_BRDCSTRCV:
2113 		*val = dmfep->rx_stats_bcast;
2114 		break;
2115 
2116 	case MAC_STAT_RBYTES:
2117 		*val = dmfep->rx_stats_rbytes;
2118 		break;
2119 
2120 	case MAC_STAT_IERRORS:
2121 		*val = dmfep->rx_stats_ierrors;
2122 		break;
2123 
2124 	case MAC_STAT_NORCVBUF:
2125 		*val = dmfep->rx_stats_norcvbuf;
2126 		break;
2127 
2128 	case MAC_STAT_COLLISIONS:
2129 		*val = dmfep->tx_stats_collisions;
2130 		break;
2131 
2132 	case MAC_STAT_OERRORS:
2133 		*val = dmfep->tx_stats_oerrors;
2134 		break;
2135 
2136 	case MAC_STAT_OPACKETS:
2137 		*val = dmfep->tx_stats_opackets;
2138 		break;
2139 
2140 	case MAC_STAT_MULTIXMT:
2141 		*val = dmfep->tx_stats_multi;
2142 		break;
2143 
2144 	case MAC_STAT_BRDCSTXMT:
2145 		*val = dmfep->tx_stats_bcast;
2146 		break;
2147 
2148 	case MAC_STAT_OBYTES:
2149 		*val = dmfep->tx_stats_obytes;
2150 		break;
2151 
2152 	case MAC_STAT_OVERFLOWS:
2153 		*val = dmfep->rx_stats_overflow;
2154 		break;
2155 
2156 	case MAC_STAT_UNDERFLOWS:
2157 		*val = dmfep->tx_stats_underflow;
2158 		break;
2159 
2160 	case ETHER_STAT_ALIGN_ERRORS:
2161 		*val = dmfep->rx_stats_align;
2162 		break;
2163 
2164 	case ETHER_STAT_FCS_ERRORS:
2165 		*val = dmfep->rx_stats_fcs;
2166 		break;
2167 
2168 	case ETHER_STAT_TOOLONG_ERRORS:
2169 		*val = dmfep->rx_stats_toolong;
2170 		break;
2171 
2172 	case ETHER_STAT_TOOSHORT_ERRORS:
2173 		*val = dmfep->rx_stats_short;
2174 		break;
2175 
2176 	case ETHER_STAT_MACRCV_ERRORS:
2177 		*val = dmfep->rx_stats_macrcv_errors;
2178 		break;
2179 
2180 	case ETHER_STAT_MACXMT_ERRORS:
2181 		*val = dmfep->tx_stats_macxmt_errors;
2182 		break;
2183 
2184 	case ETHER_STAT_JABBER_ERRORS:
2185 		*val = dmfep->tx_stats_jabber;
2186 		break;
2187 
2188 	case ETHER_STAT_CARRIER_ERRORS:
2189 		*val = dmfep->tx_stats_nocarrier;
2190 		break;
2191 
2192 	case ETHER_STAT_TX_LATE_COLLISIONS:
2193 		*val = dmfep->tx_stats_xmtlatecoll;
2194 		break;
2195 
2196 	case ETHER_STAT_EX_COLLISIONS:
2197 		*val = dmfep->tx_stats_excoll;
2198 		break;
2199 
2200 	case ETHER_STAT_DEFER_XMTS:
2201 		*val = dmfep->tx_stats_defer;
2202 		break;
2203 
2204 	case ETHER_STAT_FIRST_COLLISIONS:
2205 		*val = dmfep->tx_stats_first_coll;
2206 		break;
2207 
2208 	case ETHER_STAT_MULTI_COLLISIONS:
2209 		*val = dmfep->tx_stats_multi_coll;
2210 		break;
2211 
2212 	case ETHER_STAT_XCVR_INUSE:
2213 		*val = dmfep->phy_inuse;
2214 		break;
2215 
2216 	case ETHER_STAT_XCVR_ID:
2217 		*val = dmfep->phy_id;
2218 		break;
2219 
2220 	case ETHER_STAT_XCVR_ADDR:
2221 		*val = dmfep->phy_addr;
2222 		break;
2223 
2224 	case ETHER_STAT_LINK_DUPLEX:
2225 		*val = dmfep->op_stats_duplex;
2226 		break;
2227 
2228 	case ETHER_STAT_CAP_100T4:
2229 		*val = dmfep->param_bmsr_100T4;
2230 		break;
2231 
2232 	case ETHER_STAT_CAP_100FDX:
2233 		*val = dmfep->param_bmsr_100fdx;
2234 		break;
2235 
2236 	case ETHER_STAT_CAP_100HDX:
2237 		*val = dmfep->param_bmsr_100hdx;
2238 		break;
2239 
2240 	case ETHER_STAT_CAP_10FDX:
2241 		*val = dmfep->param_bmsr_10fdx;
2242 		break;
2243 
2244 	case ETHER_STAT_CAP_10HDX:
2245 		*val = dmfep->param_bmsr_10hdx;
2246 		break;
2247 
2248 	case ETHER_STAT_CAP_AUTONEG:
2249 		*val = dmfep->param_bmsr_autoneg;
2250 		break;
2251 
2252 	case ETHER_STAT_CAP_REMFAULT:
2253 		*val = dmfep->param_bmsr_remfault;
2254 		break;
2255 
2256 	case ETHER_STAT_ADV_CAP_AUTONEG:
2257 		*val = dmfep->param_autoneg;
2258 		break;
2259 
2260 	case ETHER_STAT_ADV_CAP_100T4:
2261 		*val = dmfep->param_anar_100T4;
2262 		break;
2263 
2264 	case ETHER_STAT_ADV_CAP_100FDX:
2265 		*val = dmfep->param_anar_100fdx;
2266 		break;
2267 
2268 	case ETHER_STAT_ADV_CAP_100HDX:
2269 		*val = dmfep->param_anar_100hdx;
2270 		break;
2271 
2272 	case ETHER_STAT_ADV_CAP_10FDX:
2273 		*val = dmfep->param_anar_10fdx;
2274 		break;
2275 
2276 	case ETHER_STAT_ADV_CAP_10HDX:
2277 		*val = dmfep->param_anar_10hdx;
2278 		break;
2279 
2280 	case ETHER_STAT_ADV_REMFAULT:
2281 		*val = dmfep->param_anar_remfault;
2282 		break;
2283 
2284 	case ETHER_STAT_LP_CAP_AUTONEG:
2285 		*val = dmfep->param_lp_autoneg;
2286 		break;
2287 
2288 	case ETHER_STAT_LP_CAP_100T4:
2289 		*val = dmfep->param_lp_100T4;
2290 		break;
2291 
2292 	case ETHER_STAT_LP_CAP_100FDX:
2293 		*val = dmfep->param_lp_100fdx;
2294 		break;
2295 
2296 	case ETHER_STAT_LP_CAP_100HDX:
2297 		*val = dmfep->param_lp_100hdx;
2298 		break;
2299 
2300 	case ETHER_STAT_LP_CAP_10FDX:
2301 		*val = dmfep->param_lp_10fdx;
2302 		break;
2303 
2304 	case ETHER_STAT_LP_CAP_10HDX:
2305 		*val = dmfep->param_lp_10hdx;
2306 		break;
2307 
2308 	case ETHER_STAT_LP_REMFAULT:
2309 		*val = dmfep->param_lp_remfault;
2310 		break;
2311 
2312 	default:
2313 		rv = ENOTSUP;
2314 	}
2315 
2316 	mutex_exit(dmfep->txlock);
2317 	mutex_exit(dmfep->rxlock);
2318 	mutex_exit(dmfep->oplock);
2319 	mutex_exit(dmfep->milock);
2320 
2321 	return (rv);
2322 }
2323 
2324 #undef	DMFE_DBG
2325 
2326 
2327 /*
2328  * ========== Ioctl handler & subfunctions ==========
2329  */
2330 
2331 #define	DMFE_DBG	DMFE_DBG_IOCTL	/* debug flag for this code	*/
2332 
2333 /*
2334  * Loopback operation
2335  *
2336  * Support access to the internal loopback and external loopback
2337  * functions selected via the Operation Mode Register (OPR).
2338  * These will be used by netlbtest (see BugId 4370609)
2339  *
2340  * Note that changing the loopback mode causes a stop/restart cycle
2341  *
2342  * It would be nice to evolve this to support the ioctls in sys/netlb.h,
2343  * but then it would be even better to use Brussels to configure this.
2344  */
2345 static enum ioc_reply
2346 dmfe_loop_ioctl(dmfe_t *dmfep, queue_t *wq, mblk_t *mp, int cmd)
2347 {
2348 	loopback_t *loop_req_p;
2349 	uint32_t loopmode;
2350 
2351 	if (mp->b_cont == NULL || MBLKL(mp->b_cont) < sizeof (loopback_t))
2352 		return (IOC_INVAL);
2353 
2354 	loop_req_p = (void *)mp->b_cont->b_rptr;
2355 
2356 	switch (cmd) {
2357 	default:
2358 		/*
2359 		 * This should never happen ...
2360 		 */
2361 		dmfe_error(dmfep, "dmfe_loop_ioctl: invalid cmd 0x%x", cmd);
2362 		return (IOC_INVAL);
2363 
2364 	case DMFE_GET_LOOP_MODE:
2365 		/*
2366 		 * This doesn't return the current loopback mode - it
2367 		 * returns a bitmask :-( of all possible loopback modes
2368 		 */
2369 		DMFE_DEBUG(("dmfe_loop_ioctl: GET_LOOP_MODE"));
2370 		loop_req_p->loopback = DMFE_LOOPBACK_MODES;
2371 		miocack(wq, mp, sizeof (loopback_t), 0);
2372 		return (IOC_DONE);
2373 
2374 	case DMFE_SET_LOOP_MODE:
2375 		/*
2376 		 * Select any of the various loopback modes
2377 		 */
2378 		DMFE_DEBUG(("dmfe_loop_ioctl: SET_LOOP_MODE %d",
2379 		    loop_req_p->loopback));
2380 		switch (loop_req_p->loopback) {
2381 		default:
2382 			return (IOC_INVAL);
2383 
2384 		case DMFE_LOOPBACK_OFF:
2385 			loopmode = LOOPBACK_OFF;
2386 			break;
2387 
2388 		case DMFE_PHY_A_LOOPBACK_ON:
2389 			loopmode = LOOPBACK_PHY_A;
2390 			break;
2391 
2392 		case DMFE_PHY_D_LOOPBACK_ON:
2393 			loopmode = LOOPBACK_PHY_D;
2394 			break;
2395 
2396 		case DMFE_INT_LOOPBACK_ON:
2397 			loopmode = LOOPBACK_INTERNAL;
2398 			break;
2399 		}
2400 
2401 		if ((dmfep->opmode & LOOPBACK_MODE_MASK) != loopmode) {
2402 			dmfep->opmode &= ~LOOPBACK_MODE_MASK;
2403 			dmfep->opmode |= loopmode;
2404 			return (IOC_RESTART_ACK);
2405 		}
2406 
2407 		return (IOC_ACK);
2408 	}
2409 }
2410 
2411 /*
2412  * Specific dmfe IOCTLs, the mac module handles the generic ones.
2413  */
2414 static void
2415 dmfe_m_ioctl(void *arg, queue_t *wq, mblk_t *mp)
2416 {
2417 	dmfe_t *dmfep = arg;
2418 	struct iocblk *iocp;
2419 	enum ioc_reply status;
2420 	int cmd;
2421 
2422 	/*
2423 	 * Validate the command before bothering with the mutexen ...
2424 	 */
2425 	iocp = (void *)mp->b_rptr;
2426 	cmd = iocp->ioc_cmd;
2427 	switch (cmd) {
2428 	default:
2429 		DMFE_DEBUG(("dmfe_m_ioctl: unknown cmd 0x%x", cmd));
2430 		miocnak(wq, mp, 0, EINVAL);
2431 		return;
2432 
2433 	case DMFE_SET_LOOP_MODE:
2434 	case DMFE_GET_LOOP_MODE:
2435 	case ND_GET:
2436 	case ND_SET:
2437 		break;
2438 	}
2439 
2440 	mutex_enter(dmfep->milock);
2441 	mutex_enter(dmfep->oplock);
2442 
2443 	switch (cmd) {
2444 	default:
2445 		_NOTE(NOTREACHED)
2446 		status = IOC_INVAL;
2447 		break;
2448 
2449 	case DMFE_SET_LOOP_MODE:
2450 	case DMFE_GET_LOOP_MODE:
2451 		status = dmfe_loop_ioctl(dmfep, wq, mp, cmd);
2452 		break;
2453 
2454 	case ND_GET:
2455 	case ND_SET:
2456 		status = dmfe_nd_ioctl(dmfep, wq, mp, cmd);
2457 		break;
2458 	}
2459 
2460 	/*
2461 	 * Do we need to restart?
2462 	 */
2463 	switch (status) {
2464 	default:
2465 		break;
2466 
2467 	case IOC_RESTART_ACK:
2468 	case IOC_RESTART:
2469 		/*
2470 		 * PHY parameters changed; we need to stop, update the
2471 		 * PHY layer and restart before sending the reply or ACK
2472 		 */
2473 		dmfe_stop(dmfep);
2474 		dmfe_update_phy(dmfep);
2475 		dmfep->update_phy = B_FALSE;
2476 
2477 		/*
2478 		 * The link will now most likely go DOWN and UP, because
2479 		 * we've changed the loopback state or the link parameters
2480 		 * or autonegotiation.  So we have to check that it's
2481 		 * settled down before we restart the TX/RX processes.
2482 		 * The ioctl code will have planted some reason strings
2483 		 * to explain what's happening, so the link state change
2484 		 * messages won't be printed on the console . We wake the
2485 		 * factotum to deal with link notifications, if any ...
2486 		 */
2487 		if (dmfe_check_link(dmfep)) {
2488 			dmfe_recheck_link(dmfep, B_TRUE);
2489 			dmfe_wake_factotum(dmfep, KS_LINK_CHECK, "ioctl");
2490 		}
2491 
2492 		if (dmfep->mac_state == DMFE_MAC_STARTED)
2493 			dmfe_start(dmfep);
2494 		break;
2495 	}
2496 
2497 	/*
2498 	 * The 'reasons-for-link-change', if any, don't apply any more
2499 	 */
2500 	mutex_exit(dmfep->oplock);
2501 	mutex_exit(dmfep->milock);
2502 
2503 	/*
2504 	 * Finally, decide how to reply
2505 	 */
2506 	switch (status) {
2507 	default:
2508 		/*
2509 		 * Error, reply with a NAK and EINVAL
2510 		 */
2511 		miocnak(wq, mp, 0, EINVAL);
2512 		break;
2513 
2514 	case IOC_RESTART_ACK:
2515 	case IOC_ACK:
2516 		/*
2517 		 * OK, reply with an ACK
2518 		 */
2519 		miocack(wq, mp, 0, 0);
2520 		break;
2521 
2522 	case IOC_RESTART:
2523 	case IOC_REPLY:
2524 		/*
2525 		 * OK, send prepared reply
2526 		 */
2527 		qreply(wq, mp);
2528 		break;
2529 
2530 	case IOC_DONE:
2531 		/*
2532 		 * OK, reply already sent
2533 		 */
2534 		break;
2535 	}
2536 }
2537 
2538 #undef	DMFE_DBG
2539 
2540 
2541 /*
2542  * ========== Per-instance setup/teardown code ==========
2543  */
2544 
2545 #define	DMFE_DBG	DMFE_DBG_INIT	/* debug flag for this code	*/
2546 
2547 /*
2548  * Determine local MAC address & broadcast address for this interface
2549  */
2550 static void
2551 dmfe_find_mac_address(dmfe_t *dmfep)
2552 {
2553 	uchar_t *prop;
2554 	uint_t propsize;
2555 	int err;
2556 
2557 	/*
2558 	 * We have to find the "vendor's factory-set address".  This is
2559 	 * the value of the property "local-mac-address", as set by OBP
2560 	 * (or a .conf file!)
2561 	 *
2562 	 * If the property is not there, then we try to find the factory
2563 	 * mac address from the devices serial EEPROM.
2564 	 */
2565 	bzero(dmfep->curr_addr, sizeof (dmfep->curr_addr));
2566 	err = ddi_prop_lookup_byte_array(DDI_DEV_T_ANY, dmfep->devinfo,
2567 	    DDI_PROP_DONTPASS, localmac_propname, &prop, &propsize);
2568 	if (err == DDI_PROP_SUCCESS) {
2569 		if (propsize == ETHERADDRL)
2570 			ethaddr_copy(prop, dmfep->curr_addr);
2571 		ddi_prop_free(prop);
2572 	} else {
2573 		/* no property set... check eeprom */
2574 		dmfe_read_eeprom(dmfep, EEPROM_EN_ADDR, dmfep->curr_addr,
2575 		    ETHERADDRL);
2576 	}
2577 
2578 	DMFE_DEBUG(("dmfe_setup_mac_address: factory %s",
2579 	    ether_sprintf((void *)dmfep->curr_addr)));
2580 }
2581 
2582 static int
2583 dmfe_alloc_dma_mem(dmfe_t *dmfep, size_t memsize,
2584 	size_t setup, size_t slop, ddi_device_acc_attr_t *attr_p,
2585 	uint_t dma_flags, dma_area_t *dma_p)
2586 {
2587 	ddi_dma_cookie_t dma_cookie;
2588 	uint_t ncookies;
2589 	int err;
2590 
2591 	/*
2592 	 * Allocate handle
2593 	 */
2594 	err = ddi_dma_alloc_handle(dmfep->devinfo, &dma_attr,
2595 	    DDI_DMA_SLEEP, NULL, &dma_p->dma_hdl);
2596 	if (err != DDI_SUCCESS)
2597 		return (DDI_FAILURE);
2598 
2599 	/*
2600 	 * Allocate memory
2601 	 */
2602 	err = ddi_dma_mem_alloc(dma_p->dma_hdl, memsize + setup + slop,
2603 	    attr_p, dma_flags & (DDI_DMA_CONSISTENT | DDI_DMA_STREAMING),
2604 	    DDI_DMA_SLEEP, NULL,
2605 	    &dma_p->mem_va, &dma_p->alength, &dma_p->acc_hdl);
2606 	if (err != DDI_SUCCESS)
2607 		return (DDI_FAILURE);
2608 
2609 	/*
2610 	 * Bind the two together
2611 	 */
2612 	err = ddi_dma_addr_bind_handle(dma_p->dma_hdl, NULL,
2613 	    dma_p->mem_va, dma_p->alength, dma_flags,
2614 	    DDI_DMA_SLEEP, NULL, &dma_cookie, &ncookies);
2615 	if (err != DDI_DMA_MAPPED)
2616 		return (DDI_FAILURE);
2617 	if ((dma_p->ncookies = ncookies) != 1)
2618 		return (DDI_FAILURE);
2619 
2620 	dma_p->mem_dvma = dma_cookie.dmac_address;
2621 	if (setup > 0) {
2622 		dma_p->setup_dvma = dma_p->mem_dvma + memsize;
2623 		dma_p->setup_va = dma_p->mem_va + memsize;
2624 	} else {
2625 		dma_p->setup_dvma = 0;
2626 		dma_p->setup_va = NULL;
2627 	}
2628 
2629 	return (DDI_SUCCESS);
2630 }
2631 
2632 /*
2633  * This function allocates the transmit and receive buffers and descriptors.
2634  */
2635 static int
2636 dmfe_alloc_bufs(dmfe_t *dmfep)
2637 {
2638 	size_t memsize;
2639 	int err;
2640 
2641 	/*
2642 	 * Allocate memory & handles for TX descriptor ring
2643 	 */
2644 	memsize = dmfep->tx.n_desc * sizeof (struct tx_desc_type);
2645 	err = dmfe_alloc_dma_mem(dmfep, memsize, SETUPBUF_SIZE, DMFE_SLOP,
2646 	    &dmfe_reg_accattr, DDI_DMA_RDWR | DDI_DMA_CONSISTENT,
2647 	    &dmfep->tx_desc);
2648 	if (err != DDI_SUCCESS)
2649 		return (DDI_FAILURE);
2650 
2651 	/*
2652 	 * Allocate memory & handles for TX buffers
2653 	 */
2654 	memsize = dmfep->tx.n_desc * DMFE_BUF_SIZE;
2655 	err = dmfe_alloc_dma_mem(dmfep, memsize, 0, 0,
2656 	    &dmfe_data_accattr, DDI_DMA_WRITE | DMFE_DMA_MODE,
2657 	    &dmfep->tx_buff);
2658 	if (err != DDI_SUCCESS)
2659 		return (DDI_FAILURE);
2660 
2661 	/*
2662 	 * Allocate memory & handles for RX descriptor ring
2663 	 */
2664 	memsize = dmfep->rx.n_desc * sizeof (struct rx_desc_type);
2665 	err = dmfe_alloc_dma_mem(dmfep, memsize, 0, DMFE_SLOP,
2666 	    &dmfe_reg_accattr, DDI_DMA_RDWR | DDI_DMA_CONSISTENT,
2667 	    &dmfep->rx_desc);
2668 	if (err != DDI_SUCCESS)
2669 		return (DDI_FAILURE);
2670 
2671 	/*
2672 	 * Allocate memory & handles for RX buffers
2673 	 */
2674 	memsize = dmfep->rx.n_desc * DMFE_BUF_SIZE;
2675 	err = dmfe_alloc_dma_mem(dmfep, memsize, 0, 0,
2676 	    &dmfe_data_accattr, DDI_DMA_READ | DMFE_DMA_MODE, &dmfep->rx_buff);
2677 	if (err != DDI_SUCCESS)
2678 		return (DDI_FAILURE);
2679 
2680 	/*
2681 	 * Allocate bitmasks for tx packet type tracking
2682 	 */
2683 	dmfep->tx_mcast = kmem_zalloc(dmfep->tx.n_desc / NBBY, KM_SLEEP);
2684 	dmfep->tx_bcast = kmem_zalloc(dmfep->tx.n_desc / NBBY, KM_SLEEP);
2685 
2686 	return (DDI_SUCCESS);
2687 }
2688 
2689 static void
2690 dmfe_free_dma_mem(dma_area_t *dma_p)
2691 {
2692 	if (dma_p->dma_hdl != NULL) {
2693 		if (dma_p->ncookies) {
2694 			(void) ddi_dma_unbind_handle(dma_p->dma_hdl);
2695 			dma_p->ncookies = 0;
2696 		}
2697 		ddi_dma_free_handle(&dma_p->dma_hdl);
2698 		dma_p->dma_hdl = NULL;
2699 		dma_p->mem_dvma = 0;
2700 		dma_p->setup_dvma = 0;
2701 	}
2702 
2703 	if (dma_p->acc_hdl != NULL) {
2704 		ddi_dma_mem_free(&dma_p->acc_hdl);
2705 		dma_p->acc_hdl = NULL;
2706 		dma_p->mem_va = NULL;
2707 		dma_p->setup_va = NULL;
2708 	}
2709 }
2710 
2711 /*
2712  * This routine frees the transmit and receive buffers and descriptors.
2713  * Make sure the chip is stopped before calling it!
2714  */
2715 static void
2716 dmfe_free_bufs(dmfe_t *dmfep)
2717 {
2718 	dmfe_free_dma_mem(&dmfep->rx_buff);
2719 	dmfe_free_dma_mem(&dmfep->rx_desc);
2720 	dmfe_free_dma_mem(&dmfep->tx_buff);
2721 	dmfe_free_dma_mem(&dmfep->tx_desc);
2722 	kmem_free(dmfep->tx_mcast, dmfep->tx.n_desc / NBBY);
2723 	kmem_free(dmfep->tx_bcast, dmfep->tx.n_desc / NBBY);
2724 }
2725 
2726 static void
2727 dmfe_unattach(dmfe_t *dmfep)
2728 {
2729 	/*
2730 	 * Clean up and free all DMFE data structures
2731 	 */
2732 	if (dmfep->cycid != NULL) {
2733 		ddi_periodic_delete(dmfep->cycid);
2734 		dmfep->cycid = NULL;
2735 	}
2736 
2737 	if (dmfep->ksp_drv != NULL)
2738 		kstat_delete(dmfep->ksp_drv);
2739 	if (dmfep->progress & PROGRESS_HWINT) {
2740 		ddi_remove_intr(dmfep->devinfo, 0, dmfep->iblk);
2741 		mutex_destroy(dmfep->txlock);
2742 		mutex_destroy(dmfep->rxlock);
2743 		mutex_destroy(dmfep->oplock);
2744 	}
2745 	if (dmfep->progress & PROGRESS_SOFTINT)
2746 		ddi_remove_softintr(dmfep->factotum_id);
2747 	if (dmfep->progress & PROGRESS_BUFS)
2748 		dmfe_free_bufs(dmfep);
2749 	if (dmfep->progress & PROGRESS_REGS)
2750 		ddi_regs_map_free(&dmfep->io_handle);
2751 	if (dmfep->progress & PROGRESS_NDD)
2752 		dmfe_nd_cleanup(dmfep);
2753 
2754 	kmem_free(dmfep, sizeof (*dmfep));
2755 }
2756 
2757 static int
2758 dmfe_config_init(dmfe_t *dmfep, chip_id_t *idp)
2759 {
2760 	ddi_acc_handle_t handle;
2761 	uint32_t regval;
2762 
2763 	if (pci_config_setup(dmfep->devinfo, &handle) != DDI_SUCCESS)
2764 		return (DDI_FAILURE);
2765 
2766 	/*
2767 	 * Get vendor/device/revision.  We expect (but don't check) that
2768 	 * (vendorid == DAVICOM_VENDOR_ID) && (deviceid == DEVICE_ID_9102)
2769 	 */
2770 	idp->vendor = pci_config_get16(handle, PCI_CONF_VENID);
2771 	idp->device = pci_config_get16(handle, PCI_CONF_DEVID);
2772 	idp->revision = pci_config_get8(handle, PCI_CONF_REVID);
2773 
2774 	/*
2775 	 * Turn on Bus Master Enable bit and ensure the device is not asleep
2776 	 */
2777 	regval = pci_config_get32(handle, PCI_CONF_COMM);
2778 	pci_config_put32(handle, PCI_CONF_COMM, (regval | PCI_COMM_ME));
2779 
2780 	regval = pci_config_get32(handle, PCI_DMFE_CONF_CFDD);
2781 	pci_config_put32(handle, PCI_DMFE_CONF_CFDD,
2782 	    regval & ~(CFDD_SLEEP | CFDD_SNOOZE));
2783 
2784 	pci_config_teardown(&handle);
2785 	return (DDI_SUCCESS);
2786 }
2787 
2788 struct ks_index {
2789 	int index;
2790 	char *name;
2791 };
2792 
2793 static const struct ks_index ks_drv_names[] = {
2794 	{	KS_INTERRUPT,			"intr"			},
2795 	{	KS_CYCLIC_RUN,			"cyclic_run"		},
2796 
2797 	{	KS_TICK_LINK_STATE,		"link_state_change"	},
2798 	{	KS_TICK_LINK_POLL,		"link_state_poll"	},
2799 	{	KS_TX_STALL,			"tx_stall_detect"	},
2800 	{	KS_CHIP_ERROR,			"chip_error_interrupt"	},
2801 
2802 	{	KS_FACTOTUM_RUN,		"factotum_run"		},
2803 	{	KS_RECOVERY,			"factotum_recover"	},
2804 	{	KS_LINK_CHECK,			"factotum_link_check"	},
2805 
2806 	{	KS_LINK_UP_CNT,			"link_up_cnt"		},
2807 	{	KS_LINK_DROP_CNT,		"link_drop_cnt"		},
2808 
2809 	{	KS_MIIREG_BMSR,			"mii_status"		},
2810 	{	KS_MIIREG_ANAR,			"mii_advert_cap"	},
2811 	{	KS_MIIREG_ANLPAR,		"mii_partner_cap"	},
2812 	{	KS_MIIREG_ANER,			"mii_expansion_cap"	},
2813 	{	KS_MIIREG_DSCSR,		"mii_dscsr"		},
2814 
2815 	{	-1,				NULL			}
2816 };
2817 
2818 static void
2819 dmfe_init_kstats(dmfe_t *dmfep, int instance)
2820 {
2821 	kstat_t *ksp;
2822 	kstat_named_t *knp;
2823 	const struct ks_index *ksip;
2824 
2825 	/* no need to create MII stats, the mac module already does it */
2826 
2827 	/* Create and initialise driver-defined kstats */
2828 	ksp = kstat_create(DRIVER_NAME, instance, "dmfe_events", "net",
2829 	    KSTAT_TYPE_NAMED, KS_DRV_COUNT, KSTAT_FLAG_PERSISTENT);
2830 	if (ksp != NULL) {
2831 		for (knp = ksp->ks_data, ksip = ks_drv_names;
2832 		    ksip->name != NULL; ++ksip) {
2833 			kstat_named_init(&knp[ksip->index], ksip->name,
2834 			    KSTAT_DATA_UINT64);
2835 		}
2836 		dmfep->ksp_drv = ksp;
2837 		dmfep->knp_drv = knp;
2838 		kstat_install(ksp);
2839 	} else {
2840 		dmfe_error(dmfep, "kstat_create() for dmfe_events failed");
2841 	}
2842 }
2843 
2844 static int
2845 dmfe_resume(dev_info_t *devinfo)
2846 {
2847 	dmfe_t *dmfep;				/* Our private data	*/
2848 	chip_id_t chipid;
2849 
2850 	dmfep = ddi_get_driver_private(devinfo);
2851 	if (dmfep == NULL)
2852 		return (DDI_FAILURE);
2853 
2854 	/*
2855 	 * Refuse to resume if the data structures aren't consistent
2856 	 */
2857 	if (dmfep->devinfo != devinfo)
2858 		return (DDI_FAILURE);
2859 
2860 	/*
2861 	 * Refuse to resume if the chip's changed its identity (*boggle*)
2862 	 */
2863 	if (dmfe_config_init(dmfep, &chipid) != DDI_SUCCESS)
2864 		return (DDI_FAILURE);
2865 	if (chipid.vendor != dmfep->chipid.vendor)
2866 		return (DDI_FAILURE);
2867 	if (chipid.device != dmfep->chipid.device)
2868 		return (DDI_FAILURE);
2869 	if (chipid.revision != dmfep->chipid.revision)
2870 		return (DDI_FAILURE);
2871 
2872 	/*
2873 	 * All OK, reinitialise h/w & kick off MAC scheduling
2874 	 */
2875 	mutex_enter(dmfep->oplock);
2876 	dmfe_restart(dmfep);
2877 	mutex_exit(dmfep->oplock);
2878 	mac_tx_update(dmfep->mh);
2879 	return (DDI_SUCCESS);
2880 }
2881 
2882 /*
2883  * attach(9E) -- Attach a device to the system
2884  *
2885  * Called once for each board successfully probed.
2886  */
2887 static int
2888 dmfe_attach(dev_info_t *devinfo, ddi_attach_cmd_t cmd)
2889 {
2890 	mac_register_t *macp;
2891 	dmfe_t *dmfep;				/* Our private data	*/
2892 	uint32_t csr6;
2893 	int instance;
2894 	int err;
2895 
2896 	instance = ddi_get_instance(devinfo);
2897 
2898 	switch (cmd) {
2899 	default:
2900 		return (DDI_FAILURE);
2901 
2902 	case DDI_RESUME:
2903 		return (dmfe_resume(devinfo));
2904 
2905 	case DDI_ATTACH:
2906 		break;
2907 	}
2908 
2909 	dmfep = kmem_zalloc(sizeof (*dmfep), KM_SLEEP);
2910 	ddi_set_driver_private(devinfo, dmfep);
2911 	dmfep->devinfo = devinfo;
2912 	dmfep->dmfe_guard = DMFE_GUARD;
2913 
2914 	/*
2915 	 * Initialize more fields in DMFE private data
2916 	 * Determine the local MAC address
2917 	 */
2918 #if	DMFEDEBUG
2919 	dmfep->debug = ddi_prop_get_int(DDI_DEV_T_ANY, devinfo, 0,
2920 	    debug_propname, dmfe_debug);
2921 #endif	/* DMFEDEBUG */
2922 	dmfep->cycid = NULL;
2923 	(void) snprintf(dmfep->ifname, sizeof (dmfep->ifname), "dmfe%d",
2924 	    instance);
2925 
2926 	/*
2927 	 * Check for custom "opmode-reg-value" property;
2928 	 * if none, use the defaults below for CSR6 ...
2929 	 */
2930 	csr6 = TX_THRESHOLD_HI | STORE_AND_FORWARD | EXT_MII_IF | OPN_25_MB1;
2931 	dmfep->opmode = ddi_prop_get_int(DDI_DEV_T_ANY, devinfo,
2932 	    DDI_PROP_DONTPASS, opmode_propname, csr6);
2933 
2934 	/*
2935 	 * Read chip ID & set up config space command register(s)
2936 	 */
2937 	if (dmfe_config_init(dmfep, &dmfep->chipid) != DDI_SUCCESS) {
2938 		dmfe_error(dmfep, "dmfe_config_init() failed");
2939 		goto attach_fail;
2940 	}
2941 	dmfep->progress |= PROGRESS_CONFIG;
2942 
2943 	/*
2944 	 * Register NDD-tweakable parameters
2945 	 */
2946 	if (dmfe_nd_init(dmfep)) {
2947 		dmfe_error(dmfep, "dmfe_nd_init() failed");
2948 		goto attach_fail;
2949 	}
2950 	dmfep->progress |= PROGRESS_NDD;
2951 
2952 	/*
2953 	 * Map operating registers
2954 	 */
2955 	err = ddi_regs_map_setup(devinfo, DMFE_PCI_RNUMBER,
2956 	    &dmfep->io_reg, 0, 0, &dmfe_reg_accattr, &dmfep->io_handle);
2957 	if (err != DDI_SUCCESS) {
2958 		dmfe_error(dmfep, "ddi_regs_map_setup() failed");
2959 		goto attach_fail;
2960 	}
2961 	dmfep->progress |= PROGRESS_REGS;
2962 
2963 	/*
2964 	 * Get our MAC address.
2965 	 */
2966 	dmfe_find_mac_address(dmfep);
2967 
2968 	/*
2969 	 * Allocate the TX and RX descriptors/buffers.
2970 	 */
2971 	dmfep->tx.n_desc = dmfe_tx_desc;
2972 	dmfep->rx.n_desc = dmfe_rx_desc;
2973 	err = dmfe_alloc_bufs(dmfep);
2974 	if (err != DDI_SUCCESS) {
2975 		dmfe_error(dmfep, "DMA buffer allocation failed");
2976 		goto attach_fail;
2977 	}
2978 	dmfep->progress |= PROGRESS_BUFS;
2979 
2980 	/*
2981 	 * Add the softint handler
2982 	 */
2983 	dmfep->link_poll_tix = factotum_start_tix;
2984 	if (ddi_add_softintr(devinfo, DDI_SOFTINT_LOW, &dmfep->factotum_id,
2985 	    NULL, NULL, dmfe_factotum, (caddr_t)dmfep) != DDI_SUCCESS) {
2986 		dmfe_error(dmfep, "ddi_add_softintr() failed");
2987 		goto attach_fail;
2988 	}
2989 	dmfep->progress |= PROGRESS_SOFTINT;
2990 
2991 	/*
2992 	 * Add the h/w interrupt handler & initialise mutexen
2993 	 */
2994 	if (ddi_add_intr(devinfo, 0, &dmfep->iblk, NULL,
2995 	    dmfe_interrupt, (caddr_t)dmfep) != DDI_SUCCESS) {
2996 		dmfe_error(dmfep, "ddi_add_intr() failed");
2997 		goto attach_fail;
2998 	}
2999 	mutex_init(dmfep->milock, NULL, MUTEX_DRIVER, NULL);
3000 	mutex_init(dmfep->oplock, NULL, MUTEX_DRIVER, dmfep->iblk);
3001 	mutex_init(dmfep->rxlock, NULL, MUTEX_DRIVER, dmfep->iblk);
3002 	mutex_init(dmfep->txlock, NULL, MUTEX_DRIVER, dmfep->iblk);
3003 	dmfep->progress |= PROGRESS_HWINT;
3004 
3005 	/*
3006 	 * Create & initialise named kstats
3007 	 */
3008 	dmfe_init_kstats(dmfep, instance);
3009 
3010 	/*
3011 	 * Reset & initialise the chip and the ring buffers
3012 	 * Initialise the (internal) PHY
3013 	 */
3014 	mutex_enter(dmfep->oplock);
3015 	mutex_enter(dmfep->rxlock);
3016 	mutex_enter(dmfep->txlock);
3017 
3018 	dmfe_reset(dmfep);
3019 
3020 	/*
3021 	 * Prepare the setup packet
3022 	 */
3023 	bzero(dmfep->tx_desc.setup_va, SETUPBUF_SIZE);
3024 	bzero(dmfep->mcast_refs, MCASTBUF_SIZE);
3025 	dmfep->addr_set = B_FALSE;
3026 	dmfep->opmode &= ~(PROMISC_MODE | PASS_MULTICAST);
3027 	dmfep->mac_state = DMFE_MAC_RESET;
3028 
3029 	mutex_exit(dmfep->txlock);
3030 	mutex_exit(dmfep->rxlock);
3031 	mutex_exit(dmfep->oplock);
3032 
3033 	dmfep->link_state = LINK_STATE_UNKNOWN;
3034 	if (dmfe_init_phy(dmfep) != B_TRUE)
3035 		goto attach_fail;
3036 	dmfep->update_phy = B_TRUE;
3037 
3038 	/*
3039 	 * Send a reasonable setup frame.  This configures our starting
3040 	 * address and the broadcast address.
3041 	 */
3042 	(void) dmfe_m_unicst(dmfep, dmfep->curr_addr);
3043 
3044 	/*
3045 	 * Initialize pointers to device specific functions which
3046 	 * will be used by the generic layer.
3047 	 */
3048 	if ((macp = mac_alloc(MAC_VERSION)) == NULL)
3049 		goto attach_fail;
3050 	macp->m_type_ident = MAC_PLUGIN_IDENT_ETHER;
3051 	macp->m_driver = dmfep;
3052 	macp->m_dip = devinfo;
3053 	macp->m_src_addr = dmfep->curr_addr;
3054 	macp->m_callbacks = &dmfe_m_callbacks;
3055 	macp->m_min_sdu = 0;
3056 	macp->m_max_sdu = ETHERMTU;
3057 	macp->m_margin = VLAN_TAGSZ;
3058 
3059 	/*
3060 	 * Finally, we're ready to register ourselves with the MAC layer
3061 	 * interface; if this succeeds, we're all ready to start()
3062 	 */
3063 	err = mac_register(macp, &dmfep->mh);
3064 	mac_free(macp);
3065 	if (err != 0)
3066 		goto attach_fail;
3067 	ASSERT(dmfep->dmfe_guard == DMFE_GUARD);
3068 
3069 	/*
3070 	 * Install the cyclic callback that we use to check for link
3071 	 * status, transmit stall, etc. The cyclic callback (dmfe_cyclic())
3072 	 * is invoked in kernel context then.
3073 	 */
3074 	ASSERT(dmfep->cycid == NULL);
3075 	dmfep->cycid = ddi_periodic_add(dmfe_cyclic, dmfep,
3076 	    dmfe_tick_us * 1000, DDI_IPL_0);
3077 	return (DDI_SUCCESS);
3078 
3079 attach_fail:
3080 	dmfe_unattach(dmfep);
3081 	return (DDI_FAILURE);
3082 }
3083 
3084 /*
3085  *	dmfe_suspend() -- suspend transmit/receive for powerdown
3086  */
3087 static int
3088 dmfe_suspend(dmfe_t *dmfep)
3089 {
3090 	/*
3091 	 * Just stop processing ...
3092 	 */
3093 	mutex_enter(dmfep->oplock);
3094 	dmfe_stop(dmfep);
3095 	mutex_exit(dmfep->oplock);
3096 
3097 	return (DDI_SUCCESS);
3098 }
3099 
3100 /*
3101  * detach(9E) -- Detach a device from the system
3102  */
3103 static int
3104 dmfe_detach(dev_info_t *devinfo, ddi_detach_cmd_t cmd)
3105 {
3106 	dmfe_t *dmfep;
3107 
3108 	dmfep = ddi_get_driver_private(devinfo);
3109 
3110 	switch (cmd) {
3111 	default:
3112 		return (DDI_FAILURE);
3113 
3114 	case DDI_SUSPEND:
3115 		return (dmfe_suspend(dmfep));
3116 
3117 	case DDI_DETACH:
3118 		break;
3119 	}
3120 
3121 	/*
3122 	 * Unregister from the MAC subsystem.  This can fail, in
3123 	 * particular if there are DLPI style-2 streams still open -
3124 	 * in which case we just return failure without shutting
3125 	 * down chip operations.
3126 	 */
3127 	if (mac_unregister(dmfep->mh) != DDI_SUCCESS)
3128 		return (DDI_FAILURE);
3129 
3130 	/*
3131 	 * All activity stopped, so we can clean up & exit
3132 	 */
3133 	dmfe_unattach(dmfep);
3134 	return (DDI_SUCCESS);
3135 }
3136 
3137 
3138 /*
3139  * ========== Module Loading Data & Entry Points ==========
3140  */
3141 
3142 DDI_DEFINE_STREAM_OPS(dmfe_dev_ops, nulldev, nulldev, dmfe_attach, dmfe_detach,
3143 	nodev, NULL, D_MP, NULL, ddi_quiesce_not_supported);
3144 
3145 static struct modldrv dmfe_modldrv = {
3146 	&mod_driverops,		/* Type of module.  This one is a driver */
3147 	dmfe_ident,		/* short description */
3148 	&dmfe_dev_ops		/* driver specific ops */
3149 };
3150 
3151 static struct modlinkage modlinkage = {
3152 	MODREV_1, (void *)&dmfe_modldrv, NULL
3153 };
3154 
3155 int
3156 _info(struct modinfo *modinfop)
3157 {
3158 	return (mod_info(&modlinkage, modinfop));
3159 }
3160 
3161 int
3162 _init(void)
3163 {
3164 	uint32_t tmp100;
3165 	uint32_t tmp10;
3166 	int i;
3167 	int status;
3168 
3169 	/* Calculate global timing parameters */
3170 	tmp100 = (dmfe_tx100_stall_us+dmfe_tick_us-1)/dmfe_tick_us;
3171 	tmp10 = (dmfe_tx10_stall_us+dmfe_tick_us-1)/dmfe_tick_us;
3172 
3173 	for (i = 0; i <= TX_PROCESS_MAX_STATE; ++i) {
3174 		switch (i) {
3175 		case TX_PROCESS_STATE(TX_PROCESS_FETCH_DATA):
3176 		case TX_PROCESS_STATE(TX_PROCESS_WAIT_END):
3177 			/*
3178 			 * The chip doesn't spontaneously recover from
3179 			 * a stall in these states, so we reset early
3180 			 */
3181 			stall_100_tix[i] = tmp100;
3182 			stall_10_tix[i] = tmp10;
3183 			break;
3184 
3185 		case TX_PROCESS_STATE(TX_PROCESS_SUSPEND):
3186 		default:
3187 			/*
3188 			 * The chip has been seen to spontaneously recover
3189 			 * after an apparent stall in the SUSPEND state,
3190 			 * so we'll allow it rather longer to do so.  As
3191 			 * stalls in other states have not been observed,
3192 			 * we'll use long timeouts for them too ...
3193 			 */
3194 			stall_100_tix[i] = tmp100 * 20;
3195 			stall_10_tix[i] = tmp10 * 20;
3196 			break;
3197 		}
3198 	}
3199 
3200 	factotum_tix = (dmfe_link_poll_us+dmfe_tick_us-1)/dmfe_tick_us;
3201 	factotum_fast_tix = 1+(factotum_tix/5);
3202 	factotum_start_tix = 1+(factotum_tix*2);
3203 
3204 	mac_init_ops(&dmfe_dev_ops, "dmfe");
3205 	status = mod_install(&modlinkage);
3206 	if (status == DDI_SUCCESS)
3207 		dmfe_log_init();
3208 
3209 	return (status);
3210 }
3211 
3212 int
3213 _fini(void)
3214 {
3215 	int status;
3216 
3217 	status = mod_remove(&modlinkage);
3218 	if (status == DDI_SUCCESS) {
3219 		mac_fini_ops(&dmfe_dev_ops);
3220 		dmfe_log_fini();
3221 	}
3222 
3223 	return (status);
3224 }
3225 
3226 #undef	DMFE_DBG
3227