xref: /titanic_41/usr/src/uts/common/io/e1000g/e1000g_tx.c (revision 2e59129a8dc96d4082395c338ad696e29471d4e0)
1 /*
2  * This file is provided under a CDDLv1 license.  When using or
3  * redistributing this file, you may do so under this license.
4  * In redistributing this file this license must be included
5  * and no other modification of this header file is permitted.
6  *
7  * CDDL LICENSE SUMMARY
8  *
9  * Copyright(c) 1999 - 2007 Intel Corporation. All rights reserved.
10  *
11  * The contents of this file are subject to the terms of Version
12  * 1.0 of the Common Development and Distribution License (the "License").
13  *
14  * You should have received a copy of the License with this software.
15  * You can obtain a copy of the License at
16  *	http://www.opensolaris.org/os/licensing.
17  * See the License for the specific language governing permissions
18  * and limitations under the License.
19  */
20 
21 /*
22  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms of the CDDLv1.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * **********************************************************************
30  *									*
31  * Module Name:								*
32  *   e1000g_tx.c							*
33  *									*
34  * Abstract:								*
35  *   This file contains some routines that take care of Transmit,	*
36  *   make the hardware to send the data pointed by the packet out	*
37  *   on to the physical medium.						*
38  *									*
39  * **********************************************************************
40  */
41 
42 #include "e1000g_sw.h"
43 #include "e1000g_debug.h"
44 
45 static boolean_t e1000g_send(struct e1000g *, mblk_t *);
46 static int e1000g_tx_copy(e1000g_tx_ring_t *,
47     p_tx_sw_packet_t, mblk_t *, uint32_t);
48 static int e1000g_tx_bind(e1000g_tx_ring_t *,
49     p_tx_sw_packet_t, mblk_t *);
50 static boolean_t check_cksum_context(e1000g_tx_ring_t *, cksum_data_t *);
51 static int e1000g_fill_tx_ring(e1000g_tx_ring_t *, LIST_DESCRIBER *,
52     cksum_data_t *);
53 static void e1000g_fill_context_descriptor(cksum_data_t *,
54     struct e1000_context_desc *);
55 static int e1000g_fill_tx_desc(e1000g_tx_ring_t *,
56     p_tx_sw_packet_t, uint64_t, size_t);
57 static uint32_t e1000g_fill_82544_desc(uint64_t Address, size_t Length,
58     p_desc_array_t desc_array);
59 static int e1000g_tx_workaround_PCIX_82544(p_tx_sw_packet_t, uint64_t, size_t);
60 static int e1000g_tx_workaround_jumbo_82544(p_tx_sw_packet_t, uint64_t, size_t);
61 static void e1000g_82547_timeout(void *);
62 static void e1000g_82547_tx_move_tail(e1000g_tx_ring_t *);
63 static void e1000g_82547_tx_move_tail_work(e1000g_tx_ring_t *);
64 
65 #ifndef E1000G_DEBUG
66 #pragma inline(e1000g_tx_copy)
67 #pragma inline(e1000g_tx_bind)
68 #pragma inline(check_cksum_context)
69 #pragma inline(e1000g_fill_tx_ring)
70 #pragma inline(e1000g_fill_context_descriptor)
71 #pragma inline(e1000g_fill_tx_desc)
72 #pragma inline(e1000g_fill_82544_desc)
73 #pragma inline(e1000g_tx_workaround_PCIX_82544)
74 #pragma inline(e1000g_tx_workaround_jumbo_82544)
75 #pragma inline(e1000g_free_tx_swpkt)
76 #endif
77 
78 /*
79  * e1000g_free_tx_swpkt	- free up the tx sw packet
80  *
81  * Unbind the previously bound DMA handle for a given
82  * transmit sw packet. And reset the sw packet data.
83  */
84 void
85 e1000g_free_tx_swpkt(register p_tx_sw_packet_t packet)
86 {
87 	switch (packet->data_transfer_type) {
88 	case USE_BCOPY:
89 		packet->tx_buf->len = 0;
90 		break;
91 #ifdef __sparc
92 	case USE_DVMA:
93 		dvma_unload(packet->tx_dma_handle, 0, -1);
94 		break;
95 #endif
96 	case USE_DMA:
97 		ddi_dma_unbind_handle(packet->tx_dma_handle);
98 		break;
99 	default:
100 		break;
101 	}
102 
103 	/*
104 	 * The mblk has been stripped off the sw packet
105 	 * and will be freed in a triggered soft intr.
106 	 */
107 	ASSERT(packet->mp == NULL);
108 
109 	packet->data_transfer_type = USE_NONE;
110 	packet->num_mblk_frag = 0;
111 	packet->num_desc = 0;
112 }
113 
114 #pragma inline(e1000g_tx_freemsg)
115 
116 void
117 e1000g_tx_freemsg(e1000g_tx_ring_t *tx_ring)
118 {
119 	mblk_t *mp;
120 
121 	if (mutex_tryenter(&tx_ring->mblks_lock) == 0)
122 		return;
123 
124 	mp = tx_ring->mblks.head;
125 
126 	tx_ring->mblks.head = NULL;
127 	tx_ring->mblks.tail = NULL;
128 
129 	mutex_exit(&tx_ring->mblks_lock);
130 
131 	if (mp != NULL)
132 		freemsgchain(mp);
133 }
134 
135 uint_t
136 e1000g_tx_softint_worker(caddr_t arg1, caddr_t arg2)
137 {
138 	struct e1000g *Adapter;
139 	mblk_t *mp;
140 
141 	Adapter = (struct e1000g *)arg1;
142 
143 	if (Adapter == NULL)
144 		return (DDI_INTR_UNCLAIMED);
145 
146 	e1000g_tx_freemsg(Adapter->tx_ring);
147 
148 	return (DDI_INTR_CLAIMED);
149 }
150 
151 mblk_t *
152 e1000g_m_tx(void *arg, mblk_t *mp)
153 {
154 	struct e1000g *Adapter = (struct e1000g *)arg;
155 	mblk_t *next;
156 
157 	rw_enter(&Adapter->chip_lock, RW_READER);
158 
159 	if (!Adapter->started || (Adapter->link_state != LINK_STATE_UP)) {
160 		freemsgchain(mp);
161 		mp = NULL;
162 	}
163 
164 	while (mp != NULL) {
165 		next = mp->b_next;
166 		mp->b_next = NULL;
167 
168 		if (!e1000g_send(Adapter, mp)) {
169 			mp->b_next = next;
170 			break;
171 		}
172 
173 		mp = next;
174 	}
175 
176 	rw_exit(&Adapter->chip_lock);
177 	return (mp);
178 }
179 
180 /*
181  * e1000g_send -  send packets onto the wire
182  *
183  * Called from e1000g_m_tx with an mblk ready to send. this
184  * routine sets up the transmit descriptors and sends data to
185  * the wire. It also pushes the just transmitted packet to
186  * the used tx sw packet list.
187  */
188 static boolean_t
189 e1000g_send(struct e1000g *Adapter, mblk_t *mp)
190 {
191 	struct e1000_hw *hw;
192 	p_tx_sw_packet_t packet;
193 	LIST_DESCRIBER pending_list;
194 	size_t len;
195 	size_t msg_size;
196 	uint32_t frag_count;
197 	int desc_count;
198 	uint32_t desc_total;
199 	uint32_t force_bcopy;
200 	mblk_t *nmp;
201 	mblk_t *tmp;
202 	e1000g_tx_ring_t *tx_ring;
203 	cksum_data_t cksum;
204 
205 	hw = &Adapter->shared;
206 	tx_ring = Adapter->tx_ring;
207 
208 	/* Get the total size and frags number of the message */
209 	force_bcopy = 0;
210 	frag_count = 0;
211 	msg_size = 0;
212 	for (nmp = mp; nmp; nmp = nmp->b_cont) {
213 		frag_count++;
214 		msg_size += MBLKL(nmp);
215 	}
216 
217 	/* Empty packet */
218 	if (msg_size == 0) {
219 		freemsg(mp);
220 		return (B_TRUE);
221 	}
222 
223 	/* Make sure packet is less than the max frame size */
224 	if (msg_size > hw->mac.max_frame_size + VLAN_TAGSZ) {
225 		/*
226 		 * For the over size packet, we'll just drop it.
227 		 * So we return B_TRUE here.
228 		 */
229 		E1000G_DEBUGLOG_1(Adapter, E1000G_WARN_LEVEL,
230 		    "Tx packet out of bound. length = %d \n", msg_size);
231 		E1000G_STAT(tx_ring->stat_over_size);
232 		freemsg(mp);
233 		return (B_TRUE);
234 	}
235 
236 	/*
237 	 * Check and reclaim tx descriptors.
238 	 * This low water mark check should be done all the time as
239 	 * Transmit interrupt delay can produce Transmit interrupts little
240 	 * late and that may cause few problems related to reaping Tx
241 	 * Descriptors... As you may run short of them before getting any
242 	 * transmit interrupt...
243 	 */
244 	if ((Adapter->tx_desc_num - tx_ring->tbd_avail) >
245 	    tx_ring->recycle_low_water) {
246 		E1000G_DEBUG_STAT(tx_ring->stat_recycle);
247 		(void) e1000g_recycle(tx_ring);
248 	}
249 
250 	if (tx_ring->tbd_avail < MAX_TX_DESC_PER_PACKET) {
251 		E1000G_DEBUG_STAT(tx_ring->stat_lack_desc);
252 		goto tx_no_resource;
253 	}
254 
255 	/*
256 	 * If there are many frags of the message, then bcopy them
257 	 * into one tx descriptor buffer will get better performance.
258 	 */
259 	if ((frag_count >= tx_ring->frags_limit) &&
260 	    (msg_size <= Adapter->tx_buffer_size)) {
261 		E1000G_DEBUG_STAT(tx_ring->stat_exceed_frags);
262 		force_bcopy |= FORCE_BCOPY_EXCEED_FRAGS;
263 	}
264 
265 	/*
266 	 * If the message size is less than the minimum ethernet packet size,
267 	 * we'll use bcopy to send it, and padd it to 60 bytes later.
268 	 */
269 	if (msg_size < MINIMUM_ETHERNET_PACKET_SIZE) {
270 		E1000G_DEBUG_STAT(tx_ring->stat_under_size);
271 		force_bcopy |= FORCE_BCOPY_UNDER_SIZE;
272 	}
273 
274 	/* Initialize variables */
275 	desc_count = 1;	/* The initial value should be greater than 0 */
276 	desc_total = 0;
277 	QUEUE_INIT_LIST(&pending_list);
278 
279 	/* Retrieve checksum info */
280 	hcksum_retrieve(mp, NULL, NULL, &cksum.cksum_start, &cksum.cksum_stuff,
281 	    NULL, NULL, &cksum.cksum_flags);
282 
283 	if (((struct ether_vlan_header *)mp->b_rptr)->ether_tpid ==
284 	    htons(ETHERTYPE_VLAN))
285 		cksum.ether_header_size = sizeof (struct ether_vlan_header);
286 	else
287 		cksum.ether_header_size = sizeof (struct ether_header);
288 
289 	/* Process each mblk fragment and fill tx descriptors */
290 	packet = NULL;
291 	nmp = mp;
292 	while (nmp) {
293 		tmp = nmp->b_cont;
294 
295 		len = MBLKL(nmp);
296 		/* Check zero length mblks */
297 		if (len == 0) {
298 			E1000G_DEBUG_STAT(tx_ring->stat_empty_frags);
299 			/*
300 			 * If there're no packet buffers have been used,
301 			 * or we just completed processing a buffer, then
302 			 * skip the empty mblk fragment.
303 			 * Otherwise, there's still a pending buffer that
304 			 * needs to be processed (tx_copy).
305 			 */
306 			if (desc_count > 0) {
307 				nmp = tmp;
308 				continue;
309 			}
310 		}
311 
312 		/*
313 		 * Get a new TxSwPacket to process mblk buffers.
314 		 */
315 		if (desc_count > 0) {
316 
317 			mutex_enter(&tx_ring->freelist_lock);
318 			packet = (p_tx_sw_packet_t)
319 			    QUEUE_POP_HEAD(&tx_ring->free_list);
320 			mutex_exit(&tx_ring->freelist_lock);
321 
322 			if (packet == NULL) {
323 				E1000G_DEBUGLOG_0(Adapter, E1000G_INFO_LEVEL,
324 				    "No Tx SwPacket available\n");
325 				E1000G_STAT(tx_ring->stat_no_swpkt);
326 				goto tx_send_failed;
327 			}
328 			QUEUE_PUSH_TAIL(&pending_list, &packet->Link);
329 		}
330 
331 		ASSERT(packet);
332 		/*
333 		 * If the size of the fragment is less than the tx_bcopy_thresh
334 		 * we'll use bcopy; Otherwise, we'll use DMA binding.
335 		 */
336 		if ((len <= Adapter->tx_bcopy_thresh) || force_bcopy) {
337 			desc_count =
338 			    e1000g_tx_copy(tx_ring, packet, nmp, force_bcopy);
339 			E1000G_DEBUG_STAT(tx_ring->stat_copy);
340 		} else {
341 			desc_count =
342 			    e1000g_tx_bind(tx_ring, packet, nmp);
343 			E1000G_DEBUG_STAT(tx_ring->stat_bind);
344 		}
345 
346 		if (desc_count > 0)
347 			desc_total += desc_count;
348 		else if (desc_count < 0)
349 			goto tx_send_failed;
350 
351 		nmp = tmp;
352 	}
353 
354 	/* Assign the message to the last sw packet */
355 	ASSERT(packet);
356 	ASSERT(packet->mp == NULL);
357 	packet->mp = mp;
358 
359 	/* Try to recycle the tx descriptors again */
360 	if (tx_ring->tbd_avail < (desc_total + 2)) {
361 		E1000G_DEBUG_STAT(tx_ring->stat_recycle_retry);
362 		(void) e1000g_recycle(tx_ring);
363 	}
364 
365 	mutex_enter(&tx_ring->tx_lock);
366 
367 	/*
368 	 * If the number of available tx descriptors is not enough for transmit
369 	 * (one redundant descriptor and one hw checksum context descriptor are
370 	 * included), then return failure.
371 	 */
372 	if (tx_ring->tbd_avail < (desc_total + 2)) {
373 		E1000G_DEBUGLOG_0(Adapter, E1000G_INFO_LEVEL,
374 		    "No Enough Tx descriptors\n");
375 		E1000G_STAT(tx_ring->stat_no_desc);
376 		mutex_exit(&tx_ring->tx_lock);
377 		goto tx_send_failed;
378 	}
379 
380 	desc_count = e1000g_fill_tx_ring(tx_ring, &pending_list, &cksum);
381 
382 	mutex_exit(&tx_ring->tx_lock);
383 
384 	ASSERT(desc_count > 0);
385 
386 	/* Send successful */
387 	return (B_TRUE);
388 
389 tx_send_failed:
390 	/* Free pending TxSwPackets */
391 	packet = (p_tx_sw_packet_t)QUEUE_GET_HEAD(&pending_list);
392 	while (packet) {
393 		packet->mp = NULL;
394 		e1000g_free_tx_swpkt(packet);
395 		packet = (p_tx_sw_packet_t)
396 		    QUEUE_GET_NEXT(&pending_list, &packet->Link);
397 	}
398 
399 	/* Return pending TxSwPackets to the "Free" list */
400 	mutex_enter(&tx_ring->freelist_lock);
401 	QUEUE_APPEND(&tx_ring->free_list, &pending_list);
402 	mutex_exit(&tx_ring->freelist_lock);
403 
404 	E1000G_STAT(tx_ring->stat_send_fail);
405 
406 	freemsg(mp);
407 
408 	/* Send failed, message dropped */
409 	return (B_TRUE);
410 
411 tx_no_resource:
412 	/*
413 	 * Enable Transmit interrupts, so that the interrupt routine can
414 	 * call mac_tx_update() when transmit descriptors become available.
415 	 */
416 	tx_ring->resched_needed = B_TRUE;
417 	if (!Adapter->tx_intr_enable)
418 		e1000g_mask_tx_interrupt(Adapter);
419 
420 	/* Message will be scheduled for re-transmit */
421 	return (B_FALSE);
422 }
423 
424 static boolean_t
425 check_cksum_context(e1000g_tx_ring_t *tx_ring, cksum_data_t *cksum)
426 {
427 	boolean_t cksum_load;
428 	cksum_data_t *last;
429 
430 	cksum_load = B_FALSE;
431 	last = &tx_ring->cksum_data;
432 
433 	if (cksum->cksum_flags != 0) {
434 		if ((cksum->ether_header_size != last->ether_header_size) ||
435 		    (cksum->cksum_flags != last->cksum_flags) ||
436 		    (cksum->cksum_stuff != last->cksum_stuff) ||
437 		    (cksum->cksum_start != last->cksum_start)) {
438 
439 			cksum_load = B_TRUE;
440 		}
441 	}
442 
443 	return (cksum_load);
444 }
445 
446 static int
447 e1000g_fill_tx_ring(e1000g_tx_ring_t *tx_ring, LIST_DESCRIBER *pending_list,
448     cksum_data_t *cksum)
449 {
450 	struct e1000g *Adapter;
451 	struct e1000_hw *hw;
452 	p_tx_sw_packet_t first_packet;
453 	p_tx_sw_packet_t packet;
454 	boolean_t cksum_load;
455 	struct e1000_tx_desc *first_data_desc;
456 	struct e1000_tx_desc *next_desc;
457 	struct e1000_tx_desc *descriptor;
458 	int desc_count;
459 	int i;
460 
461 	Adapter = tx_ring->adapter;
462 	hw = &Adapter->shared;
463 
464 	desc_count = 0;
465 	first_packet = NULL;
466 	first_data_desc = NULL;
467 	descriptor = NULL;
468 
469 	next_desc = tx_ring->tbd_next;
470 
471 	/* IP Head/TCP/UDP checksum offload */
472 	cksum_load = check_cksum_context(tx_ring, cksum);
473 
474 	if (cksum_load) {
475 		first_packet = (p_tx_sw_packet_t)QUEUE_GET_HEAD(pending_list);
476 
477 		descriptor = next_desc;
478 
479 		e1000g_fill_context_descriptor(cksum,
480 		    (struct e1000_context_desc *)descriptor);
481 
482 		/* Check the wrap-around case */
483 		if (descriptor == tx_ring->tbd_last)
484 			next_desc = tx_ring->tbd_first;
485 		else
486 			next_desc++;
487 
488 		desc_count++;
489 	}
490 
491 	first_data_desc = next_desc;
492 
493 	packet = (p_tx_sw_packet_t)QUEUE_GET_HEAD(pending_list);
494 	while (packet) {
495 		ASSERT(packet->num_desc);
496 
497 		for (i = 0; i < packet->num_desc; i++) {
498 			ASSERT(tx_ring->tbd_avail > 0);
499 
500 			descriptor = next_desc;
501 			descriptor->buffer_addr =
502 			    packet->desc[i].address;
503 			descriptor->lower.data =
504 			    packet->desc[i].length;
505 
506 			/* Zero out status */
507 			descriptor->upper.data = 0;
508 
509 			descriptor->lower.data |=
510 			    E1000_TXD_CMD_DEXT | E1000_TXD_DTYP_D;
511 			/* must set RS on every outgoing descriptor */
512 			descriptor->lower.data |=
513 			    E1000_TXD_CMD_RS;
514 
515 			/* Check the wrap-around case */
516 			if (descriptor == tx_ring->tbd_last)
517 				next_desc = tx_ring->tbd_first;
518 			else
519 				next_desc++;
520 
521 			desc_count++;
522 		}
523 
524 		if (first_packet != NULL) {
525 			/*
526 			 * Count the checksum context descriptor for
527 			 * the first SwPacket.
528 			 */
529 			first_packet->num_desc++;
530 			first_packet = NULL;
531 		}
532 
533 		packet = (p_tx_sw_packet_t)
534 		    QUEUE_GET_NEXT(pending_list, &packet->Link);
535 	}
536 
537 	ASSERT(descriptor);
538 
539 	if (cksum->cksum_flags) {
540 		if (cksum->cksum_flags & HCK_IPV4_HDRCKSUM)
541 			((struct e1000_data_desc *)first_data_desc)->
542 			    upper.fields.popts |= E1000_TXD_POPTS_IXSM;
543 		if (cksum->cksum_flags & HCK_PARTIALCKSUM)
544 			((struct e1000_data_desc *)first_data_desc)->
545 			    upper.fields.popts |= E1000_TXD_POPTS_TXSM;
546 	}
547 
548 	/*
549 	 * Last Descriptor of Packet needs End Of Packet (EOP), Report
550 	 * Status (RS) and append Ethernet CRC (IFCS) bits set.
551 	 */
552 	if (Adapter->tx_intr_delay) {
553 		descriptor->lower.data |= E1000_TXD_CMD_IDE |
554 		    E1000_TXD_CMD_EOP | E1000_TXD_CMD_IFCS;
555 	} else {
556 		descriptor->lower.data |=
557 		    E1000_TXD_CMD_EOP | E1000_TXD_CMD_IFCS;
558 	}
559 
560 	/*
561 	 * Sync the Tx descriptors DMA buffer
562 	 */
563 	(void) ddi_dma_sync(tx_ring->tbd_dma_handle,
564 	    0, 0, DDI_DMA_SYNC_FORDEV);
565 
566 	tx_ring->tbd_next = next_desc;
567 
568 	/*
569 	 * Advance the Transmit Descriptor Tail (Tdt), this tells the
570 	 * FX1000 that this frame is available to transmit.
571 	 */
572 	if (hw->mac.type == e1000_82547)
573 		e1000g_82547_tx_move_tail(tx_ring);
574 	else
575 		E1000_WRITE_REG(hw, E1000_TDT,
576 		    (uint32_t)(next_desc - tx_ring->tbd_first));
577 
578 	/* Put the pending SwPackets to the "Used" list */
579 	mutex_enter(&tx_ring->usedlist_lock);
580 	QUEUE_APPEND(&tx_ring->used_list, pending_list);
581 	tx_ring->tbd_avail -= desc_count;
582 	mutex_exit(&tx_ring->usedlist_lock);
583 
584 	/* Store the cksum data */
585 	if (cksum_load)
586 		tx_ring->cksum_data = *cksum;
587 
588 	return (desc_count);
589 }
590 
591 
592 /*
593  * e1000g_tx_setup - setup tx data structures
594  *
595  * This routine initializes all of the transmit related
596  * structures. This includes the Transmit descriptors,
597  * and the tx_sw_packet structures.
598  */
599 void
600 e1000g_tx_setup(struct e1000g *Adapter)
601 {
602 	struct e1000_hw *hw;
603 	p_tx_sw_packet_t packet;
604 	UINT i;
605 	uint32_t buf_high;
606 	uint32_t buf_low;
607 	uint32_t reg_tipg;
608 	uint32_t reg_tctl;
609 	uint32_t reg_tarc;
610 	uint16_t speed, duplex;
611 	int size;
612 	e1000g_tx_ring_t *tx_ring;
613 
614 	hw = &Adapter->shared;
615 	tx_ring = Adapter->tx_ring;
616 
617 	/* init the lists */
618 	/*
619 	 * Here we don't need to protect the lists using the
620 	 * usedlist_lock and freelist_lock, for they have
621 	 * been protected by the chip_lock.
622 	 */
623 	QUEUE_INIT_LIST(&tx_ring->used_list);
624 	QUEUE_INIT_LIST(&tx_ring->free_list);
625 
626 	/* Go through and set up each SW_Packet */
627 	packet = tx_ring->packet_area;
628 	for (i = 0; i < Adapter->tx_freelist_num; i++, packet++) {
629 		/* Initialize this tx_sw_apcket area */
630 		e1000g_free_tx_swpkt(packet);
631 		/* Add this tx_sw_packet to the free list */
632 		QUEUE_PUSH_TAIL(&tx_ring->free_list,
633 		    &packet->Link);
634 	}
635 
636 	/* Setup TX descriptor pointers */
637 	tx_ring->tbd_next = tx_ring->tbd_first;
638 	tx_ring->tbd_oldest = tx_ring->tbd_first;
639 
640 	/*
641 	 * Setup Hardware TX Registers
642 	 */
643 	/* Setup the Transmit Control Register (TCTL). */
644 	reg_tctl = E1000_TCTL_PSP | E1000_TCTL_EN |
645 	    (E1000_COLLISION_THRESHOLD << E1000_CT_SHIFT) |
646 	    (E1000_COLLISION_DISTANCE << E1000_COLD_SHIFT) |
647 	    E1000_TCTL_RTLC;
648 
649 	/* Enable the MULR bit */
650 	if (hw->bus.type == e1000_bus_type_pci_express)
651 		reg_tctl |= E1000_TCTL_MULR;
652 
653 	E1000_WRITE_REG(hw, E1000_TCTL, reg_tctl);
654 
655 	if ((hw->mac.type == e1000_82571) || (hw->mac.type == e1000_82572)) {
656 		e1000_get_speed_and_duplex(hw, &speed, &duplex);
657 
658 		reg_tarc = E1000_READ_REG(hw, E1000_TARC0);
659 		reg_tarc |= (1 << 25);
660 		if (speed == SPEED_1000)
661 			reg_tarc |= (1 << 21);
662 		E1000_WRITE_REG(hw, E1000_TARC0, reg_tarc);
663 
664 		reg_tarc = E1000_READ_REG(hw, E1000_TARC1);
665 		reg_tarc |= (1 << 25);
666 		if (reg_tctl & E1000_TCTL_MULR)
667 			reg_tarc &= ~(1 << 28);
668 		else
669 			reg_tarc |= (1 << 28);
670 		E1000_WRITE_REG(hw, E1000_TARC1, reg_tarc);
671 
672 	} else if (hw->mac.type == e1000_80003es2lan) {
673 		reg_tarc = E1000_READ_REG(hw, E1000_TARC0);
674 		reg_tarc |= 1;
675 		if (hw->media_type == e1000_media_type_internal_serdes)
676 			reg_tarc |= (1 << 20);
677 		E1000_WRITE_REG(hw, E1000_TARC0, reg_tarc);
678 
679 		reg_tarc = E1000_READ_REG(hw, E1000_TARC1);
680 		reg_tarc |= 1;
681 		E1000_WRITE_REG(hw, E1000_TARC1, reg_tarc);
682 	}
683 
684 	/* Setup HW Base and Length of Tx descriptor area */
685 	size = (Adapter->tx_desc_num * sizeof (struct e1000_tx_desc));
686 	E1000_WRITE_REG(hw, E1000_TDLEN, size);
687 	size = E1000_READ_REG(hw, E1000_TDLEN);
688 
689 	buf_low = (uint32_t)tx_ring->tbd_dma_addr;
690 	buf_high = (uint32_t)(tx_ring->tbd_dma_addr >> 32);
691 
692 	E1000_WRITE_REG(hw, E1000_TDBAL, buf_low);
693 	E1000_WRITE_REG(hw, E1000_TDBAH, buf_high);
694 
695 	/* Setup our HW Tx Head & Tail descriptor pointers */
696 	E1000_WRITE_REG(hw, E1000_TDH, 0);
697 	E1000_WRITE_REG(hw, E1000_TDT, 0);
698 
699 	/* Set the default values for the Tx Inter Packet Gap timer */
700 	if ((hw->mac.type == e1000_82542) &&
701 	    ((hw->revision_id == E1000_REVISION_2) ||
702 	    (hw->revision_id == E1000_REVISION_3))) {
703 		reg_tipg = DEFAULT_82542_TIPG_IPGT;
704 		reg_tipg |=
705 		    DEFAULT_82542_TIPG_IPGR1 << E1000_TIPG_IPGR1_SHIFT;
706 		reg_tipg |=
707 		    DEFAULT_82542_TIPG_IPGR2 << E1000_TIPG_IPGR2_SHIFT;
708 	} else {
709 		if (hw->media_type == e1000_media_type_fiber)
710 			reg_tipg = DEFAULT_82543_TIPG_IPGT_FIBER;
711 		else
712 			reg_tipg = DEFAULT_82543_TIPG_IPGT_COPPER;
713 		reg_tipg |=
714 		    DEFAULT_82543_TIPG_IPGR1 << E1000_TIPG_IPGR1_SHIFT;
715 		reg_tipg |=
716 		    DEFAULT_82543_TIPG_IPGR2 << E1000_TIPG_IPGR2_SHIFT;
717 	}
718 	E1000_WRITE_REG(hw, E1000_TIPG, reg_tipg);
719 
720 	/* Setup Transmit Interrupt Delay Value */
721 	if (Adapter->tx_intr_delay) {
722 		E1000_WRITE_REG(hw, E1000_TIDV, Adapter->tx_intr_delay);
723 	}
724 
725 	tx_ring->tbd_avail = Adapter->tx_desc_num;
726 
727 	/* For TCP/UDP checksum offload */
728 	tx_ring->cksum_data.cksum_stuff = 0;
729 	tx_ring->cksum_data.cksum_start = 0;
730 	tx_ring->cksum_data.cksum_flags = 0;
731 	tx_ring->cksum_data.ether_header_size = 0;
732 }
733 
734 /*
735  * e1000g_recycle - recycle the tx descriptors and tx sw packets
736  */
737 int
738 e1000g_recycle(e1000g_tx_ring_t *tx_ring)
739 {
740 	struct e1000g *Adapter;
741 	LIST_DESCRIBER pending_list;
742 	p_tx_sw_packet_t packet;
743 	mblk_t *mp;
744 	mblk_t *nmp;
745 	struct e1000_tx_desc *descriptor;
746 	int desc_count;
747 
748 	/*
749 	 * This function will examine each TxSwPacket in the 'used' queue
750 	 * if the e1000g is done with it then the associated resources (Tx
751 	 * Descriptors) will be "freed" and the TxSwPacket will be
752 	 * returned to the 'free' queue.
753 	 */
754 	Adapter = tx_ring->adapter;
755 
756 	desc_count = 0;
757 	QUEUE_INIT_LIST(&pending_list);
758 
759 	mutex_enter(&tx_ring->usedlist_lock);
760 
761 	packet = (p_tx_sw_packet_t)QUEUE_GET_HEAD(&tx_ring->used_list);
762 	if (packet == NULL) {
763 		mutex_exit(&tx_ring->usedlist_lock);
764 		tx_ring->recycle_fail = 0;
765 		tx_ring->stall_watchdog = 0;
766 		return (0);
767 	}
768 
769 	/* Sync the Tx descriptor DMA buffer */
770 	(void) ddi_dma_sync(tx_ring->tbd_dma_handle,
771 	    0, 0, DDI_DMA_SYNC_FORKERNEL);
772 
773 	/*
774 	 * While there are still TxSwPackets in the used queue check them
775 	 */
776 	while (packet =
777 	    (p_tx_sw_packet_t)QUEUE_GET_HEAD(&tx_ring->used_list)) {
778 
779 		/*
780 		 * Get hold of the next descriptor that the e1000g will
781 		 * report status back to (this will be the last descriptor
782 		 * of a given sw packet). We only want to free the
783 		 * sw packet (and it resources) if the e1000g is done
784 		 * with ALL of the descriptors.  If the e1000g is done
785 		 * with the last one then it is done with all of them.
786 		 */
787 		ASSERT(packet->num_desc);
788 		descriptor = tx_ring->tbd_oldest + (packet->num_desc - 1);
789 
790 		/* Check for wrap case */
791 		if (descriptor > tx_ring->tbd_last)
792 			descriptor -= Adapter->tx_desc_num;
793 
794 		/*
795 		 * If the descriptor done bit is set free TxSwPacket and
796 		 * associated resources
797 		 */
798 		if (descriptor->upper.fields.status & E1000_TXD_STAT_DD) {
799 			QUEUE_POP_HEAD(&tx_ring->used_list);
800 			QUEUE_PUSH_TAIL(&pending_list, &packet->Link);
801 
802 			if (descriptor == tx_ring->tbd_last)
803 				tx_ring->tbd_oldest =
804 				    tx_ring->tbd_first;
805 			else
806 				tx_ring->tbd_oldest =
807 				    descriptor + 1;
808 
809 			desc_count += packet->num_desc;
810 
811 			if (desc_count >= tx_ring->recycle_num)
812 				break;
813 		} else {
814 			/*
815 			 * Found a sw packet that the e1000g is not done
816 			 * with then there is no reason to check the rest
817 			 * of the queue.
818 			 */
819 			break;
820 		}
821 	}
822 
823 	tx_ring->tbd_avail += desc_count;
824 
825 	mutex_exit(&tx_ring->usedlist_lock);
826 
827 	if (desc_count == 0) {
828 		tx_ring->recycle_fail++;
829 		E1000G_DEBUG_STAT(tx_ring->stat_recycle_none);
830 		return (0);
831 	}
832 
833 	tx_ring->recycle_fail = 0;
834 	tx_ring->stall_watchdog = 0;
835 
836 	mp = NULL;
837 	nmp = NULL;
838 	packet = (p_tx_sw_packet_t)QUEUE_GET_HEAD(&pending_list);
839 	ASSERT(packet != NULL);
840 	while (packet != NULL) {
841 		if (packet->mp != NULL) {
842 			ASSERT(packet->mp->b_next == NULL);
843 			/* Assemble the message chain */
844 			if (mp == NULL) {
845 				mp = packet->mp;
846 				nmp = packet->mp;
847 			} else {
848 				nmp->b_next = packet->mp;
849 				nmp = packet->mp;
850 			}
851 			/* Disconnect the message from the sw packet */
852 			packet->mp = NULL;
853 		}
854 
855 		/* Free the TxSwPackets */
856 		e1000g_free_tx_swpkt(packet);
857 
858 		packet = (p_tx_sw_packet_t)
859 		    QUEUE_GET_NEXT(&pending_list, &packet->Link);
860 	}
861 
862 	/* Save the message chain */
863 	if (mp != NULL) {
864 		mutex_enter(&tx_ring->mblks_lock);
865 		if (tx_ring->mblks.head == NULL) {
866 			tx_ring->mblks.head = mp;
867 			tx_ring->mblks.tail = nmp;
868 		} else {
869 			tx_ring->mblks.tail->b_next = mp;
870 			tx_ring->mblks.tail = nmp;
871 		}
872 		mutex_exit(&tx_ring->mblks_lock);
873 
874 		/*
875 		 * If the tx interrupt is enabled, the messages will be freed
876 		 * in the tx interrupt; Otherwise, they are freed here by
877 		 * triggering a soft interrupt.
878 		 */
879 		if (!Adapter->tx_intr_enable)
880 			ddi_intr_trigger_softint(Adapter->tx_softint_handle,
881 			    NULL);
882 	}
883 
884 	/* Return the TxSwPackets back to the FreeList */
885 	mutex_enter(&tx_ring->freelist_lock);
886 	QUEUE_APPEND(&tx_ring->free_list, &pending_list);
887 	mutex_exit(&tx_ring->freelist_lock);
888 
889 	return (desc_count);
890 }
891 
892 /*
893  * 82544 Coexistence issue workaround:
894  *    There are 2 issues.
895  *    1. If a 32 bit split completion happens from P64H2 and another
896  *	agent drives a 64 bit request/split completion after ONLY
897  *	1 idle clock (BRCM/Emulex/Adaptec fiber channel cards) then
898  *	82544 has a problem where in to clock all the data in, it
899  *	looks at REQ64# signal and since it has changed so fast (i.e. 1
900  *	idle clock turn around), it will fail to clock all the data in.
901  *	Data coming from certain ending addresses has exposure to this issue.
902  *
903  * To detect this issue, following equation can be used...
904  *	SIZE[3:0] + ADDR[2:0] = SUM[3:0].
905  *	If SUM[3:0] is in between 1 to 4, we will have this issue.
906  *
907  * ROOT CAUSE:
908  *	The erratum involves the 82544 PCIX elasticity FIFO implementations as
909  *	64-bit FIFO's and flushing of the final partial-bytes corresponding
910  *	to the end of a requested read burst. Under a specific burst condition
911  *	of ending-data alignment and 32-byte split-completions, the final
912  *	byte(s) of split-completion data require an extra clock cycle to flush
913  *	into 64-bit FIFO orientation.  An incorrect logic dependency on the
914  *	REQ64# signal occurring during during this clock cycle may cause the
915  *	residual byte(s) to be lost, thereby rendering the internal DMA client
916  *	forever awaiting the final byte(s) for an outbound data-fetch.  The
917  *	erratum is confirmed to *only* occur if certain subsequent external
918  *	64-bit PCIX bus transactions occur immediately (minimum possible bus
919  *	turn- around) following the odd-aligned 32-bit split-completion
920  *	containing the final byte(s).  Intel has confirmed that this has been
921  *	seen only with chipset/bridges which have the capability to provide
922  *	32-bit split-completion data, and in the presence of newer PCIX bus
923  *	agents which fully-optimize the inter-transaction turn-around (zero
924  *	additional initiator latency when pre-granted bus ownership).
925  *
926  *   	This issue does not exist in PCI bus mode, when any agent is operating
927  *	in 32 bit only mode or on chipsets that do not do 32 bit split
928  *	completions for 64 bit read requests (Serverworks chipsets). P64H2 does
929  *	32 bit split completions for any read request that has bit 2 set to 1
930  *	for the requested address and read request size is more than 8 bytes.
931  *
932  *   2. Another issue is related to 82544 driving DACs under the similar
933  *	scenario (32 bit split completion followed by 64 bit transaction with
934  *	only 1 cycle turnaround). This issue is still being root caused. We
935  *	think that both of these issues can be avoided if following workaround
936  *	is implemented. It seems DAC issues is related to ending addresses being
937  *	0x9, 0xA, 0xB, 0xC and hence ending up at odd boundaries in elasticity
938  *	FIFO which does not get flushed due to REQ64# dependency. We will only
939  *	know the full story after it has been simulated successfully by HW team.
940  *
941  * WORKAROUND:
942  *	Make sure we do not have ending address as 1,2,3,4(Hang) or 9,a,b,c(DAC)
943  */
944 static uint32_t
945 e1000g_fill_82544_desc(uint64_t address,
946     size_t length, p_desc_array_t desc_array)
947 {
948 	/*
949 	 * Since issue is sensitive to length and address.
950 	 * Let us first check the address...
951 	 */
952 	uint32_t safe_terminator;
953 
954 	if (length <= 4) {
955 		desc_array->descriptor[0].address = address;
956 		desc_array->descriptor[0].length = length;
957 		desc_array->elements = 1;
958 		return (desc_array->elements);
959 	}
960 	safe_terminator =
961 	    (uint32_t)((((uint32_t)address & 0x7) +
962 	    (length & 0xF)) & 0xF);
963 	/*
964 	 * if it does not fall between 0x1 to 0x4 and 0x9 to 0xC then
965 	 * return
966 	 */
967 	if (safe_terminator == 0 ||
968 	    (safe_terminator > 4 && safe_terminator < 9) ||
969 	    (safe_terminator > 0xC && safe_terminator <= 0xF)) {
970 		desc_array->descriptor[0].address = address;
971 		desc_array->descriptor[0].length = length;
972 		desc_array->elements = 1;
973 		return (desc_array->elements);
974 	}
975 
976 	desc_array->descriptor[0].address = address;
977 	desc_array->descriptor[0].length = length - 4;
978 	desc_array->descriptor[1].address = address + (length - 4);
979 	desc_array->descriptor[1].length = 4;
980 	desc_array->elements = 2;
981 	return (desc_array->elements);
982 }
983 
984 static int
985 e1000g_tx_copy(e1000g_tx_ring_t *tx_ring, p_tx_sw_packet_t packet,
986     mblk_t *mp, uint32_t force_bcopy)
987 {
988 	size_t len;
989 	size_t len1;
990 	dma_buffer_t *tx_buf;
991 	mblk_t *nmp;
992 	boolean_t finished;
993 	int desc_count;
994 
995 	desc_count = 0;
996 	tx_buf = packet->tx_buf;
997 	len = MBLKL(mp);
998 
999 	ASSERT((tx_buf->len + len) <= tx_buf->size);
1000 
1001 	if (len > 0) {
1002 		bcopy(mp->b_rptr,
1003 		    tx_buf->address + tx_buf->len,
1004 		    len);
1005 		tx_buf->len += len;
1006 
1007 		packet->num_mblk_frag++;
1008 	}
1009 
1010 	nmp = mp->b_cont;
1011 	if (nmp == NULL) {
1012 		finished = B_TRUE;
1013 	} else {
1014 		len1 = MBLKL(nmp);
1015 		if ((tx_buf->len + len1) > tx_buf->size)
1016 			finished = B_TRUE;
1017 		else if (force_bcopy)
1018 			finished = B_FALSE;
1019 		else if (len1 > tx_ring->adapter->tx_bcopy_thresh)
1020 			finished = B_TRUE;
1021 		else
1022 			finished = B_FALSE;
1023 	}
1024 
1025 	if (finished) {
1026 		E1000G_DEBUG_STAT_COND(tx_ring->stat_multi_copy,
1027 		    (tx_buf->len > len));
1028 
1029 		/*
1030 		 * If the packet is smaller than 64 bytes, which is the
1031 		 * minimum ethernet packet size, pad the packet to make
1032 		 * it at least 60 bytes. The hardware will add 4 bytes
1033 		 * for CRC.
1034 		 */
1035 		if (force_bcopy & FORCE_BCOPY_UNDER_SIZE) {
1036 			ASSERT(tx_buf->len < MINIMUM_ETHERNET_PACKET_SIZE);
1037 
1038 			bzero(tx_buf->address + tx_buf->len,
1039 			    MINIMUM_ETHERNET_PACKET_SIZE - tx_buf->len);
1040 			tx_buf->len = MINIMUM_ETHERNET_PACKET_SIZE;
1041 		}
1042 
1043 #ifdef __sparc
1044 		if (packet->dma_type == USE_DVMA)
1045 			dvma_sync(tx_buf->dma_handle, 0, DDI_DMA_SYNC_FORDEV);
1046 		else
1047 			(void) ddi_dma_sync(tx_buf->dma_handle, 0,
1048 			    tx_buf->len, DDI_DMA_SYNC_FORDEV);
1049 #else
1050 		(void) ddi_dma_sync(tx_buf->dma_handle, 0,
1051 		    tx_buf->len, DDI_DMA_SYNC_FORDEV);
1052 #endif
1053 
1054 		packet->data_transfer_type = USE_BCOPY;
1055 
1056 		desc_count = e1000g_fill_tx_desc(tx_ring,
1057 		    packet,
1058 		    tx_buf->dma_address,
1059 		    tx_buf->len);
1060 
1061 		if (desc_count <= 0)
1062 			return (-1);
1063 	}
1064 
1065 	return (desc_count);
1066 }
1067 
1068 static int
1069 e1000g_tx_bind(e1000g_tx_ring_t *tx_ring, p_tx_sw_packet_t packet, mblk_t *mp)
1070 {
1071 	int j;
1072 	int mystat;
1073 	size_t len;
1074 	ddi_dma_cookie_t dma_cookie;
1075 	uint_t ncookies;
1076 	int desc_count;
1077 	uint32_t desc_total;
1078 
1079 	desc_total = 0;
1080 	len = MBLKL(mp);
1081 
1082 	/*
1083 	 * ddi_dma_addr_bind_handle() allocates  DMA  resources  for  a
1084 	 * memory  object such that a device can perform DMA to or from
1085 	 * the object.  DMA resources  are  allocated  considering  the
1086 	 * device's  DMA  attributes  as  expressed by ddi_dma_attr(9S)
1087 	 * (see ddi_dma_alloc_handle(9F)).
1088 	 *
1089 	 * ddi_dma_addr_bind_handle() fills in  the  first  DMA  cookie
1090 	 * pointed  to by cookiep with the appropriate address, length,
1091 	 * and bus type. *ccountp is set to the number of DMA  cookies
1092 	 * representing this DMA object. Subsequent DMA cookies must be
1093 	 * retrieved by calling ddi_dma_nextcookie(9F)  the  number  of
1094 	 * times specified by *countp - 1.
1095 	 */
1096 	switch (packet->dma_type) {
1097 #ifdef __sparc
1098 	case USE_DVMA:
1099 		dvma_kaddr_load(packet->tx_dma_handle,
1100 		    (caddr_t)mp->b_rptr, len, 0, &dma_cookie);
1101 
1102 		dvma_sync(packet->tx_dma_handle, 0,
1103 		    DDI_DMA_SYNC_FORDEV);
1104 
1105 		ncookies = 1;
1106 		packet->data_transfer_type = USE_DVMA;
1107 		break;
1108 #endif
1109 	case USE_DMA:
1110 		if ((mystat = ddi_dma_addr_bind_handle(
1111 		    packet->tx_dma_handle, NULL,
1112 		    (caddr_t)mp->b_rptr, len,
1113 		    DDI_DMA_WRITE | DDI_DMA_STREAMING,
1114 		    DDI_DMA_DONTWAIT, 0, &dma_cookie,
1115 		    &ncookies)) != DDI_DMA_MAPPED) {
1116 
1117 			e1000g_log(tx_ring->adapter, CE_WARN,
1118 			    "Couldn't bind mblk buffer to Tx DMA handle: "
1119 			    "return: %X, Pkt: %X\n",
1120 			    mystat, packet);
1121 			return (-1);
1122 		}
1123 
1124 		/*
1125 		 * An implicit ddi_dma_sync() is done when the
1126 		 * ddi_dma_addr_bind_handle() is called. So we
1127 		 * don't need to explicitly call ddi_dma_sync()
1128 		 * here any more.
1129 		 */
1130 		ASSERT(ncookies);
1131 		E1000G_DEBUG_STAT_COND(tx_ring->stat_multi_cookie,
1132 		    (ncookies > 1));
1133 
1134 		/*
1135 		 * The data_transfer_type value must be set after the handle
1136 		 * has been bound, for it will be used in e1000g_free_tx_swpkt()
1137 		 * to decide whether we need to unbind the handle.
1138 		 */
1139 		packet->data_transfer_type = USE_DMA;
1140 		break;
1141 	default:
1142 		ASSERT(B_FALSE);
1143 		break;
1144 	}
1145 
1146 	packet->num_mblk_frag++;
1147 
1148 	/*
1149 	 * Each address could span thru multpile cookie..
1150 	 * Each cookie will have one descriptor
1151 	 */
1152 	for (j = ncookies; j != 0; j--) {
1153 
1154 		desc_count = e1000g_fill_tx_desc(tx_ring,
1155 		    packet,
1156 		    dma_cookie.dmac_laddress,
1157 		    dma_cookie.dmac_size);
1158 
1159 		if (desc_count <= 0)
1160 			return (-1);
1161 
1162 		desc_total += desc_count;
1163 
1164 		/*
1165 		 * ddi_dma_nextcookie() retrieves subsequent DMA
1166 		 * cookies for a DMA object.
1167 		 * ddi_dma_nextcookie() fills in the
1168 		 * ddi_dma_cookie(9S) structure pointed to by
1169 		 * cookiep.  The ddi_dma_cookie(9S) structure
1170 		 * must be allocated prior to calling
1171 		 * ddi_dma_nextcookie(). The DMA cookie count
1172 		 * returned by ddi_dma_buf_bind_handle(9F),
1173 		 * ddi_dma_addr_bind_handle(9F), or
1174 		 * ddi_dma_getwin(9F) indicates the number of DMA
1175 		 * cookies a DMA object consists of.  If the
1176 		 * resulting cookie count, N, is larger than 1,
1177 		 * ddi_dma_nextcookie() must be called N-1 times
1178 		 * to retrieve all DMA cookies.
1179 		 */
1180 		if (j > 1) {
1181 			ddi_dma_nextcookie(packet->tx_dma_handle,
1182 			    &dma_cookie);
1183 		}
1184 	}
1185 
1186 	return (desc_total);
1187 }
1188 
1189 static void
1190 e1000g_fill_context_descriptor(cksum_data_t *cksum,
1191     struct e1000_context_desc *cksum_desc)
1192 {
1193 	if (cksum->cksum_flags & HCK_IPV4_HDRCKSUM) {
1194 		cksum_desc->lower_setup.ip_fields.ipcss =
1195 		    cksum->ether_header_size;
1196 		cksum_desc->lower_setup.ip_fields.ipcso =
1197 		    cksum->ether_header_size +
1198 		    offsetof(struct ip, ip_sum);
1199 		cksum_desc->lower_setup.ip_fields.ipcse =
1200 		    cksum->ether_header_size +
1201 		    sizeof (struct ip) - 1;
1202 	} else
1203 		cksum_desc->lower_setup.ip_config = 0;
1204 
1205 	if (cksum->cksum_flags & HCK_PARTIALCKSUM) {
1206 		/*
1207 		 * The packet with same protocol has the following
1208 		 * stuff and start offset:
1209 		 * |  Protocol  | Stuff  | Start  | Checksum
1210 		 * |		| Offset | Offset | Enable
1211 		 * | IPv4 + TCP |  0x24  |  0x14  |  Yes
1212 		 * | IPv4 + UDP |  0x1A  |  0x14  |  Yes
1213 		 * | IPv6 + TCP |  0x20  |  0x10  |  No
1214 		 * | IPv6 + UDP |  0x14  |  0x10  |  No
1215 		 */
1216 		cksum_desc->upper_setup.tcp_fields.tucss =
1217 		    cksum->cksum_start + cksum->ether_header_size;
1218 		cksum_desc->upper_setup.tcp_fields.tucso =
1219 		    cksum->cksum_stuff + cksum->ether_header_size;
1220 		cksum_desc->upper_setup.tcp_fields.tucse = 0;
1221 	} else
1222 		cksum_desc->upper_setup.tcp_config = 0;
1223 
1224 	cksum_desc->cmd_and_length = E1000_TXD_CMD_DEXT;
1225 
1226 	/*
1227 	 * Zero out the options for TCP Segmentation Offload,
1228 	 * since we don't support it in this version
1229 	 */
1230 	cksum_desc->tcp_seg_setup.data = 0;
1231 }
1232 
1233 static int
1234 e1000g_fill_tx_desc(e1000g_tx_ring_t *tx_ring,
1235     p_tx_sw_packet_t packet, uint64_t address, size_t size)
1236 {
1237 	struct e1000_hw *hw = &tx_ring->adapter->shared;
1238 	p_sw_desc_t desc;
1239 
1240 	if (hw->mac.type == e1000_82544) {
1241 		if (hw->bus.type == e1000_bus_type_pcix)
1242 			return (e1000g_tx_workaround_PCIX_82544(packet,
1243 			    address, size));
1244 
1245 		if (size > JUMBO_FRAG_LENGTH)
1246 			return (e1000g_tx_workaround_jumbo_82544(packet,
1247 			    address, size));
1248 	}
1249 
1250 	ASSERT(packet->num_desc < MAX_TX_DESC_PER_PACKET);
1251 
1252 	desc = &packet->desc[packet->num_desc];
1253 	desc->address = address;
1254 	desc->length = size;
1255 
1256 	packet->num_desc++;
1257 
1258 	return (1);
1259 }
1260 
1261 static int
1262 e1000g_tx_workaround_PCIX_82544(p_tx_sw_packet_t packet,
1263     uint64_t address, size_t size)
1264 {
1265 	p_sw_desc_t desc;
1266 	int desc_count;
1267 	long size_left;
1268 	size_t len;
1269 	uint32_t counter;
1270 	uint32_t array_elements;
1271 	desc_array_t desc_array;
1272 
1273 	/*
1274 	 * Coexist Workaround for cordova: RP: 07/04/03
1275 	 *
1276 	 * RP: ERRATA: Workaround ISSUE:
1277 	 * 8kb_buffer_Lockup CONTROLLER: Cordova Breakup
1278 	 * Eachbuffer in to 8kb pieces until the
1279 	 * remainder is < 8kb
1280 	 */
1281 	size_left = size;
1282 	desc_count = 0;
1283 
1284 	while (size_left > 0) {
1285 		if (size_left > MAX_TX_BUF_SIZE)
1286 			len = MAX_TX_BUF_SIZE;
1287 		else
1288 			len = size_left;
1289 
1290 		array_elements = e1000g_fill_82544_desc(address,
1291 		    len, &desc_array);
1292 
1293 		for (counter = 0; counter < array_elements; counter++) {
1294 			ASSERT(packet->num_desc < MAX_TX_DESC_PER_PACKET);
1295 			/*
1296 			 * Put in the buffer address
1297 			 */
1298 			desc = &packet->desc[packet->num_desc];
1299 
1300 			desc->address =
1301 			    desc_array.descriptor[counter].address;
1302 			desc->length =
1303 			    desc_array.descriptor[counter].length;
1304 
1305 			packet->num_desc++;
1306 			desc_count++;
1307 		} /* for */
1308 
1309 		/*
1310 		 * Update the buffer address and length
1311 		 */
1312 		address += MAX_TX_BUF_SIZE;
1313 		size_left -= MAX_TX_BUF_SIZE;
1314 	} /* while */
1315 
1316 	return (desc_count);
1317 }
1318 
1319 static int
1320 e1000g_tx_workaround_jumbo_82544(p_tx_sw_packet_t packet,
1321     uint64_t address, size_t size)
1322 {
1323 	p_sw_desc_t desc;
1324 	int desc_count;
1325 	long size_left;
1326 	uint32_t offset;
1327 
1328 	/*
1329 	 * Workaround for Jumbo Frames on Cordova
1330 	 * PSD 06/01/2001
1331 	 */
1332 	size_left = size;
1333 	desc_count = 0;
1334 	offset = 0;
1335 	while (size_left > 0) {
1336 		ASSERT(packet->num_desc < MAX_TX_DESC_PER_PACKET);
1337 
1338 		desc = &packet->desc[packet->num_desc];
1339 
1340 		desc->address = address + offset;
1341 
1342 		if (size_left > JUMBO_FRAG_LENGTH)
1343 			desc->length = JUMBO_FRAG_LENGTH;
1344 		else
1345 			desc->length = size_left;
1346 
1347 		packet->num_desc++;
1348 		desc_count++;
1349 
1350 		offset += desc->length;
1351 		size_left -= JUMBO_FRAG_LENGTH;
1352 	}
1353 
1354 	return (desc_count);
1355 }
1356 
1357 #pragma inline(e1000g_82547_tx_move_tail_work)
1358 
1359 static void
1360 e1000g_82547_tx_move_tail_work(e1000g_tx_ring_t *tx_ring)
1361 {
1362 	struct e1000_hw *hw;
1363 	uint16_t hw_tdt;
1364 	uint16_t sw_tdt;
1365 	struct e1000_tx_desc *tx_desc;
1366 	uint16_t length = 0;
1367 	boolean_t eop = B_FALSE;
1368 	struct e1000g *Adapter;
1369 
1370 	Adapter = tx_ring->adapter;
1371 	hw = &Adapter->shared;
1372 
1373 	hw_tdt = E1000_READ_REG(hw, E1000_TDT);
1374 	sw_tdt = tx_ring->tbd_next - tx_ring->tbd_first;
1375 
1376 	while (hw_tdt != sw_tdt) {
1377 		tx_desc = &(tx_ring->tbd_first[hw_tdt]);
1378 		length += tx_desc->lower.flags.length;
1379 		eop = tx_desc->lower.data & E1000_TXD_CMD_EOP;
1380 		if (++hw_tdt == Adapter->tx_desc_num)
1381 			hw_tdt = 0;
1382 
1383 		if (eop) {
1384 			if ((Adapter->link_duplex == HALF_DUPLEX) &&
1385 			    (e1000_fifo_workaround_82547(hw, length)
1386 			    != E1000_SUCCESS)) {
1387 				if (tx_ring->timer_enable_82547) {
1388 					ASSERT(tx_ring->timer_id_82547 == 0);
1389 					tx_ring->timer_id_82547 =
1390 					    timeout(e1000g_82547_timeout,
1391 					    (void *)tx_ring,
1392 					    drv_usectohz(10000));
1393 				}
1394 				return;
1395 
1396 			} else {
1397 				E1000_WRITE_REG(hw, E1000_TDT, hw_tdt);
1398 				e1000_update_tx_fifo_head_82547(hw, length);
1399 				length = 0;
1400 			}
1401 		}
1402 	}
1403 }
1404 
1405 static void
1406 e1000g_82547_timeout(void *arg)
1407 {
1408 	e1000g_tx_ring_t *tx_ring;
1409 
1410 	tx_ring = (e1000g_tx_ring_t *)arg;
1411 
1412 	mutex_enter(&tx_ring->tx_lock);
1413 
1414 	tx_ring->timer_id_82547 = 0;
1415 	e1000g_82547_tx_move_tail_work(tx_ring);
1416 
1417 	mutex_exit(&tx_ring->tx_lock);
1418 }
1419 
1420 static void
1421 e1000g_82547_tx_move_tail(e1000g_tx_ring_t *tx_ring)
1422 {
1423 	timeout_id_t tid;
1424 
1425 	ASSERT(MUTEX_HELD(&tx_ring->tx_lock));
1426 
1427 	tid = tx_ring->timer_id_82547;
1428 	tx_ring->timer_id_82547 = 0;
1429 	if (tid != 0) {
1430 		tx_ring->timer_enable_82547 = B_FALSE;
1431 		mutex_exit(&tx_ring->tx_lock);
1432 
1433 		(void) untimeout(tid);
1434 
1435 		mutex_enter(&tx_ring->tx_lock);
1436 	}
1437 	tx_ring->timer_enable_82547 = B_TRUE;
1438 	e1000g_82547_tx_move_tail_work(tx_ring);
1439 }
1440