xref: /titanic_41/usr/src/uts/common/io/nxge/nxge_send.c (revision 9e86db79b7d1bbc5f2f04e99954cbd5eae0e22bb)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/mac_provider.h>
27 #include <sys/nxge/nxge_impl.h>
28 #include <sys/nxge/nxge_hio.h>
29 #include <npi_tx_wr64.h>
30 
31 /* Software LSO required header files */
32 #include <netinet/tcp.h>
33 #include <inet/ip_impl.h>
34 #include <inet/tcp.h>
35 
36 extern uint64_t mac_pkt_hash(uint_t, mblk_t *mp, uint8_t policy,
37     boolean_t is_outbound);
38 
39 static mblk_t *nxge_lso_eliminate(mblk_t *);
40 static mblk_t *nxge_do_softlso(mblk_t *mp, uint32_t mss);
41 static void nxge_lso_info_get(mblk_t *, uint32_t *, uint32_t *);
42 static void nxge_hcksum_retrieve(mblk_t *,
43     uint32_t *, uint32_t *, uint32_t *,
44     uint32_t *, uint32_t *);
45 static uint32_t nxge_csgen(uint16_t *, int);
46 
47 extern uint32_t		nxge_reclaim_pending;
48 extern uint32_t 	nxge_bcopy_thresh;
49 extern uint32_t 	nxge_dvma_thresh;
50 extern uint32_t 	nxge_dma_stream_thresh;
51 extern uint32_t		nxge_tx_minfree;
52 extern uint32_t		nxge_tx_intr_thres;
53 extern uint32_t		nxge_tx_max_gathers;
54 extern uint32_t		nxge_tx_tiny_pack;
55 extern uint32_t		nxge_tx_use_bcopy;
56 extern nxge_tx_mode_t	nxge_tx_scheme;
57 uint32_t		nxge_lso_kick_cnt = 2;
58 
59 
60 void
61 nxge_tx_ring_task(void *arg)
62 {
63 	p_tx_ring_t	ring = (p_tx_ring_t)arg;
64 
65 	MUTEX_ENTER(&ring->lock);
66 	(void) nxge_txdma_reclaim(ring->nxgep, ring, 0);
67 	MUTEX_EXIT(&ring->lock);
68 
69 	if (!ring->tx_ring_offline) {
70 		mac_tx_ring_update(ring->nxgep->mach, ring->tx_ring_handle);
71 	}
72 }
73 
74 static void
75 nxge_tx_ring_dispatch(p_tx_ring_t ring)
76 {
77 	/*
78 	 * Kick the ring task to reclaim some buffers.
79 	 */
80 	(void) ddi_taskq_dispatch(ring->taskq,
81 	    nxge_tx_ring_task, (void *)ring, DDI_SLEEP);
82 }
83 
84 mblk_t *
85 nxge_tx_ring_send(void *arg, mblk_t *mp)
86 {
87 	p_nxge_ring_handle_t	nrhp = (p_nxge_ring_handle_t)arg;
88 	p_nxge_t		nxgep;
89 	p_tx_ring_t		tx_ring_p;
90 	int			status, channel;
91 
92 	ASSERT(nrhp != NULL);
93 	nxgep = nrhp->nxgep;
94 	channel = nxgep->pt_config.hw_config.tdc.start + nrhp->index;
95 	tx_ring_p = nxgep->tx_rings->rings[channel];
96 
97 	/*
98 	 * We may be in a transition from offlined DMA to onlined
99 	 * DMA.
100 	 */
101 	if (tx_ring_p == NULL) {
102 		ASSERT(tx_ring_p != NULL);
103 		freemsg(mp);
104 		return ((mblk_t *)NULL);
105 	}
106 
107 	/*
108 	 * Valid DMA?
109 	 */
110 	ASSERT(nxgep == tx_ring_p->nxgep);
111 
112 	/*
113 	 * Make sure DMA is not offlined.
114 	 */
115 	if (isLDOMservice(nxgep) && tx_ring_p->tx_ring_offline) {
116 		ASSERT(!tx_ring_p->tx_ring_offline);
117 		freemsg(mp);
118 		return ((mblk_t *)NULL);
119 	}
120 
121 	/*
122 	 * Transmit the packet.
123 	 */
124 	status = nxge_start(nxgep, tx_ring_p, mp);
125 	if (status) {
126 		nxge_tx_ring_dispatch(tx_ring_p);
127 		return (mp);
128 	}
129 
130 	return ((mblk_t *)NULL);
131 }
132 
133 int
134 nxge_start(p_nxge_t nxgep, p_tx_ring_t tx_ring_p, p_mblk_t mp)
135 {
136 	int 			dma_status, status = 0;
137 	p_tx_desc_t 		tx_desc_ring_vp;
138 	npi_handle_t		npi_desc_handle;
139 	nxge_os_dma_handle_t 	tx_desc_dma_handle;
140 	p_tx_desc_t 		tx_desc_p;
141 	p_tx_msg_t 		tx_msg_ring;
142 	p_tx_msg_t 		tx_msg_p;
143 	tx_desc_t		tx_desc, *tmp_desc_p;
144 	tx_desc_t		sop_tx_desc, *sop_tx_desc_p;
145 	p_tx_pkt_header_t	hdrp;
146 	tx_pkt_hdr_all_t	tmp_hdrp;
147 	p_tx_pkt_hdr_all_t	pkthdrp;
148 	uint8_t			npads = 0;
149 	uint64_t 		dma_ioaddr;
150 	uint32_t		dma_flags;
151 	int			last_bidx;
152 	uint8_t 		*b_rptr;
153 	caddr_t 		kaddr;
154 	uint32_t		nmblks;
155 	uint32_t		ngathers;
156 	uint32_t		clen;
157 	int 			len;
158 	uint32_t		pkt_len, pack_len, min_len;
159 	uint32_t		bcopy_thresh;
160 	int 			i, cur_index, sop_index;
161 	uint16_t		tail_index;
162 	boolean_t		tail_wrap = B_FALSE;
163 	nxge_dma_common_t	desc_area;
164 	nxge_os_dma_handle_t 	dma_handle;
165 	ddi_dma_cookie_t 	dma_cookie;
166 	npi_handle_t		npi_handle;
167 	p_mblk_t 		nmp;
168 	p_mblk_t		t_mp;
169 	uint32_t 		ncookies;
170 	boolean_t 		good_packet;
171 	boolean_t 		mark_mode = B_FALSE;
172 	p_nxge_stats_t 		statsp;
173 	p_nxge_tx_ring_stats_t tdc_stats;
174 	t_uscalar_t 		start_offset = 0;
175 	t_uscalar_t 		stuff_offset = 0;
176 	t_uscalar_t 		end_offset = 0;
177 	t_uscalar_t 		value = 0;
178 	t_uscalar_t 		cksum_flags = 0;
179 	boolean_t		cksum_on = B_FALSE;
180 	uint32_t		boff = 0;
181 	uint64_t		tot_xfer_len = 0;
182 	boolean_t		header_set = B_FALSE;
183 #ifdef NXGE_DEBUG
184 	p_tx_desc_t 		tx_desc_ring_pp;
185 	p_tx_desc_t 		tx_desc_pp;
186 	tx_desc_t		*save_desc_p;
187 	int			dump_len;
188 	int			sad_len;
189 	uint64_t		sad;
190 	int			xfer_len;
191 	uint32_t		msgsize;
192 #endif
193 	p_mblk_t 		mp_chain = NULL;
194 	boolean_t		is_lso = B_FALSE;
195 	boolean_t		lso_again;
196 	int			cur_index_lso;
197 	p_mblk_t 		nmp_lso_save;
198 	uint32_t		lso_ngathers;
199 	boolean_t		lso_tail_wrap = B_FALSE;
200 
201 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
202 	    "==> nxge_start: tx dma channel %d", tx_ring_p->tdc));
203 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
204 	    "==> nxge_start: Starting tdc %d desc pending %d",
205 	    tx_ring_p->tdc, tx_ring_p->descs_pending));
206 
207 	statsp = nxgep->statsp;
208 
209 	if (!isLDOMguest(nxgep)) {
210 		switch (nxgep->mac.portmode) {
211 		default:
212 			if (nxgep->statsp->port_stats.lb_mode ==
213 			    nxge_lb_normal) {
214 				if (!statsp->mac_stats.link_up) {
215 					freemsg(mp);
216 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
217 					    "==> nxge_start: "
218 					    "link not up"));
219 					goto nxge_start_fail1;
220 				}
221 			}
222 			break;
223 		case PORT_10G_FIBER:
224 			/*
225 			 * For the following modes, check the link status
226 			 * before sending the packet out:
227 			 * nxge_lb_normal,
228 			 * nxge_lb_ext10g,
229 			 * nxge_lb_ext1000,
230 			 * nxge_lb_ext100,
231 			 * nxge_lb_ext10.
232 			 */
233 			if (nxgep->statsp->port_stats.lb_mode <
234 			    nxge_lb_phy10g) {
235 				if (!statsp->mac_stats.link_up) {
236 					freemsg(mp);
237 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
238 					    "==> nxge_start: "
239 					    "link not up"));
240 					goto nxge_start_fail1;
241 				}
242 			}
243 			break;
244 		}
245 	}
246 
247 	if ((!(nxgep->drv_state & STATE_HW_INITIALIZED)) ||
248 	    (nxgep->nxge_mac_state != NXGE_MAC_STARTED)) {
249 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
250 		    "==> nxge_start: hardware not initialized or stopped"));
251 		freemsg(mp);
252 		goto nxge_start_fail1;
253 	}
254 
255 	if (nxgep->soft_lso_enable) {
256 		mp_chain = nxge_lso_eliminate(mp);
257 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
258 		    "==> nxge_start(0): LSO mp $%p mp_chain $%p",
259 		    mp, mp_chain));
260 		if (mp_chain == NULL) {
261 			NXGE_ERROR_MSG((nxgep, TX_CTL,
262 			    "==> nxge_send(0): NULL mp_chain $%p != mp $%p",
263 			    mp_chain, mp));
264 			goto nxge_start_fail1;
265 		}
266 		if (mp_chain != mp) {
267 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
268 			    "==> nxge_send(1): IS LSO mp_chain $%p != mp $%p",
269 			    mp_chain, mp));
270 			is_lso = B_TRUE;
271 			mp = mp_chain;
272 			mp_chain = mp_chain->b_next;
273 			mp->b_next = NULL;
274 		}
275 	}
276 
277 	hcksum_retrieve(mp, NULL, NULL, &start_offset,
278 	    &stuff_offset, &end_offset, &value, &cksum_flags);
279 	if (!NXGE_IS_VLAN_PACKET(mp->b_rptr)) {
280 		start_offset += sizeof (ether_header_t);
281 		stuff_offset += sizeof (ether_header_t);
282 	} else {
283 		start_offset += sizeof (struct ether_vlan_header);
284 		stuff_offset += sizeof (struct ether_vlan_header);
285 	}
286 
287 	if (cksum_flags & HCK_PARTIALCKSUM) {
288 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
289 		    "==> nxge_start: mp $%p len %d "
290 		    "cksum_flags 0x%x (partial checksum) ",
291 		    mp, MBLKL(mp), cksum_flags));
292 		cksum_on = B_TRUE;
293 	}
294 
295 	pkthdrp = (p_tx_pkt_hdr_all_t)&tmp_hdrp;
296 	pkthdrp->reserved = 0;
297 	tmp_hdrp.pkthdr.value = 0;
298 	nxge_fill_tx_hdr(mp, B_FALSE, cksum_on,
299 	    0, 0, pkthdrp,
300 	    start_offset, stuff_offset);
301 
302 	lso_again = B_FALSE;
303 	lso_ngathers = 0;
304 
305 	MUTEX_ENTER(&tx_ring_p->lock);
306 
307 	if (isLDOMservice(nxgep)) {
308 		tx_ring_p->tx_ring_busy = B_TRUE;
309 		if (tx_ring_p->tx_ring_offline) {
310 			freemsg(mp);
311 			tx_ring_p->tx_ring_busy = B_FALSE;
312 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
313 			    NXGE_TX_RING_OFFLINED);
314 			MUTEX_EXIT(&tx_ring_p->lock);
315 			return (status);
316 		}
317 	}
318 
319 	cur_index_lso = tx_ring_p->wr_index;
320 	lso_tail_wrap = tx_ring_p->wr_index_wrap;
321 start_again:
322 	ngathers = 0;
323 	sop_index = tx_ring_p->wr_index;
324 #ifdef	NXGE_DEBUG
325 	if (tx_ring_p->descs_pending) {
326 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
327 		    "desc pending %d ", tx_ring_p->descs_pending));
328 	}
329 
330 	dump_len = (int)(MBLKL(mp));
331 	dump_len = (dump_len > 128) ? 128: dump_len;
332 
333 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
334 	    "==> nxge_start: tdc %d: dumping ...: b_rptr $%p "
335 	    "(Before header reserve: ORIGINAL LEN %d)",
336 	    tx_ring_p->tdc,
337 	    mp->b_rptr,
338 	    dump_len));
339 
340 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: dump packets "
341 	    "(IP ORIGINAL b_rptr $%p): %s", mp->b_rptr,
342 	    nxge_dump_packet((char *)mp->b_rptr, dump_len)));
343 #endif
344 
345 	tdc_stats = tx_ring_p->tdc_stats;
346 	mark_mode = (tx_ring_p->descs_pending &&
347 	    (((int)tx_ring_p->tx_ring_size - (int)tx_ring_p->descs_pending) <
348 	    (int)nxge_tx_minfree));
349 
350 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
351 	    "TX Descriptor ring is channel %d mark mode %d",
352 	    tx_ring_p->tdc, mark_mode));
353 
354 	if ((tx_ring_p->descs_pending + lso_ngathers) >= nxge_reclaim_pending) {
355 		if (!nxge_txdma_reclaim(nxgep, tx_ring_p,
356 		    (nxge_tx_minfree + lso_ngathers))) {
357 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
358 			    "TX Descriptor ring is full: channel %d",
359 			    tx_ring_p->tdc));
360 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
361 			    "TX Descriptor ring is full: channel %d",
362 			    tx_ring_p->tdc));
363 			if (is_lso) {
364 				/*
365 				 * free the current mp and mp_chain if not FULL.
366 				 */
367 				tdc_stats->tx_no_desc++;
368 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
369 				    "LSO packet: TX Descriptor ring is full: "
370 				    "channel %d",
371 				    tx_ring_p->tdc));
372 				goto nxge_start_fail_lso;
373 			} else {
374 				(void) cas32((uint32_t *)&tx_ring_p->queueing,
375 				    0, 1);
376 				tdc_stats->tx_no_desc++;
377 
378 				if (isLDOMservice(nxgep)) {
379 					tx_ring_p->tx_ring_busy = B_FALSE;
380 					if (tx_ring_p->tx_ring_offline) {
381 						(void) atomic_swap_32(
382 						    &tx_ring_p->tx_ring_offline,
383 						    NXGE_TX_RING_OFFLINED);
384 					}
385 				}
386 
387 				MUTEX_EXIT(&tx_ring_p->lock);
388 				status = 1;
389 				goto nxge_start_fail1;
390 			}
391 		}
392 	}
393 
394 	nmp = mp;
395 	i = sop_index = tx_ring_p->wr_index;
396 	nmblks = 0;
397 	ngathers = 0;
398 	pkt_len = 0;
399 	pack_len = 0;
400 	clen = 0;
401 	last_bidx = -1;
402 	good_packet = B_TRUE;
403 
404 	desc_area = tx_ring_p->tdc_desc;
405 	npi_handle = desc_area.npi_handle;
406 	npi_desc_handle.regh = (nxge_os_acc_handle_t)
407 	    DMA_COMMON_ACC_HANDLE(desc_area);
408 	tx_desc_ring_vp = (p_tx_desc_t)DMA_COMMON_VPTR(desc_area);
409 	tx_desc_dma_handle = (nxge_os_dma_handle_t)
410 	    DMA_COMMON_HANDLE(desc_area);
411 	tx_msg_ring = tx_ring_p->tx_msg_ring;
412 
413 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: wr_index %d i %d",
414 	    sop_index, i));
415 
416 #ifdef	NXGE_DEBUG
417 	msgsize = msgdsize(nmp);
418 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
419 	    "==> nxge_start(1): wr_index %d i %d msgdsize %d",
420 	    sop_index, i, msgsize));
421 #endif
422 	/*
423 	 * The first 16 bytes of the premapped buffer are reserved
424 	 * for header. No padding will be used.
425 	 */
426 	pkt_len = pack_len = boff = TX_PKT_HEADER_SIZE;
427 	if (nxge_tx_use_bcopy && (nxgep->niu_type != N2_NIU)) {
428 		bcopy_thresh = (nxge_bcopy_thresh - TX_PKT_HEADER_SIZE);
429 	} else {
430 		bcopy_thresh = (TX_BCOPY_SIZE - TX_PKT_HEADER_SIZE);
431 	}
432 	while (nmp) {
433 		good_packet = B_TRUE;
434 		b_rptr = nmp->b_rptr;
435 		len = MBLKL(nmp);
436 		if (len <= 0) {
437 			nmp = nmp->b_cont;
438 			continue;
439 		}
440 		nmblks++;
441 
442 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(1): nmblks %d "
443 		    "len %d pkt_len %d pack_len %d",
444 		    nmblks, len, pkt_len, pack_len));
445 		/*
446 		 * Hardware limits the transfer length to 4K for NIU and
447 		 * 4076 (TX_MAX_TRANSFER_LENGTH) for Neptune. But we just
448 		 * use TX_MAX_TRANSFER_LENGTH as the limit for both.
449 		 * If len is longer than the limit, then we break nmp into
450 		 * two chunks: Make the first chunk equal to the limit and
451 		 * the second chunk for the remaining data. If the second
452 		 * chunk is still larger than the limit, then it will be
453 		 * broken into two in the next pass.
454 		 */
455 		if (len > TX_MAX_TRANSFER_LENGTH - TX_PKT_HEADER_SIZE) {
456 			if ((t_mp = dupb(nmp)) != NULL) {
457 				nmp->b_wptr = nmp->b_rptr +
458 				    (TX_MAX_TRANSFER_LENGTH
459 				    - TX_PKT_HEADER_SIZE);
460 				t_mp->b_rptr = nmp->b_wptr;
461 				t_mp->b_cont = nmp->b_cont;
462 				nmp->b_cont = t_mp;
463 				len = MBLKL(nmp);
464 			} else {
465 				if (is_lso) {
466 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
467 					    "LSO packet: dupb failed: "
468 					    "channel %d",
469 					    tx_ring_p->tdc));
470 					mp = nmp;
471 					goto nxge_start_fail_lso;
472 				} else {
473 					good_packet = B_FALSE;
474 					goto nxge_start_fail2;
475 				}
476 			}
477 		}
478 		tx_desc.value = 0;
479 		tx_desc_p = &tx_desc_ring_vp[i];
480 #ifdef	NXGE_DEBUG
481 		tx_desc_pp = &tx_desc_ring_pp[i];
482 #endif
483 		tx_msg_p = &tx_msg_ring[i];
484 #if defined(__i386)
485 		npi_desc_handle.regp = (uint32_t)tx_desc_p;
486 #else
487 		npi_desc_handle.regp = (uint64_t)tx_desc_p;
488 #endif
489 		if (!header_set &&
490 		    ((!nxge_tx_use_bcopy && (len > TX_BCOPY_SIZE)) ||
491 		    (len >= bcopy_thresh))) {
492 			header_set = B_TRUE;
493 			bcopy_thresh += TX_PKT_HEADER_SIZE;
494 			boff = 0;
495 			pack_len = 0;
496 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
497 			hdrp = (p_tx_pkt_header_t)kaddr;
498 			clen = pkt_len;
499 			dma_handle = tx_msg_p->buf_dma_handle;
500 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
501 			(void) ddi_dma_sync(dma_handle,
502 			    i * nxge_bcopy_thresh, nxge_bcopy_thresh,
503 			    DDI_DMA_SYNC_FORDEV);
504 
505 			tx_msg_p->flags.dma_type = USE_BCOPY;
506 			goto nxge_start_control_header_only;
507 		}
508 
509 		pkt_len += len;
510 		pack_len += len;
511 
512 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(3): "
513 		    "desc entry %d "
514 		    "DESC IOADDR $%p "
515 		    "desc_vp $%p tx_desc_p $%p "
516 		    "desc_pp $%p tx_desc_pp $%p "
517 		    "len %d pkt_len %d pack_len %d",
518 		    i,
519 		    DMA_COMMON_IOADDR(desc_area),
520 		    tx_desc_ring_vp, tx_desc_p,
521 		    tx_desc_ring_pp, tx_desc_pp,
522 		    len, pkt_len, pack_len));
523 
524 		if (len < bcopy_thresh) {
525 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(4): "
526 			    "USE BCOPY: "));
527 			if (nxge_tx_tiny_pack) {
528 				uint32_t blst =
529 				    TXDMA_DESC_NEXT_INDEX(i, -1,
530 				    tx_ring_p->tx_wrap_mask);
531 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
532 				    "==> nxge_start(5): pack"));
533 				if ((pack_len <= bcopy_thresh) &&
534 				    (last_bidx == blst)) {
535 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
536 					    "==> nxge_start: pack(6) "
537 					    "(pkt_len %d pack_len %d)",
538 					    pkt_len, pack_len));
539 					i = blst;
540 					tx_desc_p = &tx_desc_ring_vp[i];
541 #ifdef	NXGE_DEBUG
542 					tx_desc_pp = &tx_desc_ring_pp[i];
543 #endif
544 					tx_msg_p = &tx_msg_ring[i];
545 					boff = pack_len - len;
546 					ngathers--;
547 				} else if (pack_len > bcopy_thresh &&
548 				    header_set) {
549 					pack_len = len;
550 					boff = 0;
551 					bcopy_thresh = nxge_bcopy_thresh;
552 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
553 					    "==> nxge_start(7): > max NEW "
554 					    "bcopy thresh %d "
555 					    "pkt_len %d pack_len %d(next)",
556 					    bcopy_thresh,
557 					    pkt_len, pack_len));
558 				}
559 				last_bidx = i;
560 			}
561 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
562 			if ((boff == TX_PKT_HEADER_SIZE) && (nmblks == 1)) {
563 				hdrp = (p_tx_pkt_header_t)kaddr;
564 				header_set = B_TRUE;
565 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
566 				    "==> nxge_start(7_x2): "
567 				    "pkt_len %d pack_len %d (new hdrp $%p)",
568 				    pkt_len, pack_len, hdrp));
569 			}
570 			tx_msg_p->flags.dma_type = USE_BCOPY;
571 			kaddr += boff;
572 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(8): "
573 			    "USE BCOPY: before bcopy "
574 			    "DESC IOADDR $%p entry %d "
575 			    "bcopy packets %d "
576 			    "bcopy kaddr $%p "
577 			    "bcopy ioaddr (SAD) $%p "
578 			    "bcopy clen %d "
579 			    "bcopy boff %d",
580 			    DMA_COMMON_IOADDR(desc_area), i,
581 			    tdc_stats->tx_hdr_pkts,
582 			    kaddr,
583 			    dma_ioaddr,
584 			    clen,
585 			    boff));
586 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
587 			    "1USE BCOPY: "));
588 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
589 			    "2USE BCOPY: "));
590 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
591 			    "last USE BCOPY: copy from b_rptr $%p "
592 			    "to KADDR $%p (len %d offset %d",
593 			    b_rptr, kaddr, len, boff));
594 
595 			bcopy(b_rptr, kaddr, len);
596 
597 #ifdef	NXGE_DEBUG
598 			dump_len = (len > 128) ? 128: len;
599 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
600 			    "==> nxge_start: dump packets "
601 			    "(After BCOPY len %d)"
602 			    "(b_rptr $%p): %s", len, nmp->b_rptr,
603 			    nxge_dump_packet((char *)nmp->b_rptr,
604 			    dump_len)));
605 #endif
606 
607 			dma_handle = tx_msg_p->buf_dma_handle;
608 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
609 			(void) ddi_dma_sync(dma_handle,
610 			    i * nxge_bcopy_thresh, nxge_bcopy_thresh,
611 			    DDI_DMA_SYNC_FORDEV);
612 			clen = len + boff;
613 			tdc_stats->tx_hdr_pkts++;
614 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(9): "
615 			    "USE BCOPY: "
616 			    "DESC IOADDR $%p entry %d "
617 			    "bcopy packets %d "
618 			    "bcopy kaddr $%p "
619 			    "bcopy ioaddr (SAD) $%p "
620 			    "bcopy clen %d "
621 			    "bcopy boff %d",
622 			    DMA_COMMON_IOADDR(desc_area),
623 			    i,
624 			    tdc_stats->tx_hdr_pkts,
625 			    kaddr,
626 			    dma_ioaddr,
627 			    clen,
628 			    boff));
629 		} else {
630 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(12): "
631 			    "USE DVMA: len %d", len));
632 			tx_msg_p->flags.dma_type = USE_DMA;
633 			dma_flags = DDI_DMA_WRITE;
634 			if (len < nxge_dma_stream_thresh) {
635 				dma_flags |= DDI_DMA_CONSISTENT;
636 			} else {
637 				dma_flags |= DDI_DMA_STREAMING;
638 			}
639 
640 			dma_handle = tx_msg_p->dma_handle;
641 			dma_status = ddi_dma_addr_bind_handle(dma_handle, NULL,
642 			    (caddr_t)b_rptr, len, dma_flags,
643 			    DDI_DMA_DONTWAIT, NULL,
644 			    &dma_cookie, &ncookies);
645 			if (dma_status == DDI_DMA_MAPPED) {
646 				dma_ioaddr = dma_cookie.dmac_laddress;
647 				len = (int)dma_cookie.dmac_size;
648 				clen = (uint32_t)dma_cookie.dmac_size;
649 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
650 				    "==> nxge_start(12_1): "
651 				    "USE DVMA: len %d clen %d "
652 				    "ngathers %d",
653 				    len, clen,
654 				    ngathers));
655 #if defined(__i386)
656 				npi_desc_handle.regp = (uint32_t)tx_desc_p;
657 #else
658 				npi_desc_handle.regp = (uint64_t)tx_desc_p;
659 #endif
660 				while (ncookies > 1) {
661 					ngathers++;
662 					/*
663 					 * this is the fix for multiple
664 					 * cookies, which are basically
665 					 * a descriptor entry, we don't set
666 					 * SOP bit as well as related fields
667 					 */
668 
669 					(void) npi_txdma_desc_gather_set(
670 					    npi_desc_handle,
671 					    &tx_desc,
672 					    (ngathers -1),
673 					    mark_mode,
674 					    ngathers,
675 					    dma_ioaddr,
676 					    clen);
677 
678 					tx_msg_p->tx_msg_size = clen;
679 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
680 					    "==> nxge_start:  DMA "
681 					    "ncookie %d "
682 					    "ngathers %d "
683 					    "dma_ioaddr $%p len %d"
684 					    "desc $%p descp $%p (%d)",
685 					    ncookies,
686 					    ngathers,
687 					    dma_ioaddr, clen,
688 					    *tx_desc_p, tx_desc_p, i));
689 
690 					ddi_dma_nextcookie(dma_handle,
691 					    &dma_cookie);
692 					dma_ioaddr =
693 					    dma_cookie.dmac_laddress;
694 
695 					len = (int)dma_cookie.dmac_size;
696 					clen = (uint32_t)dma_cookie.dmac_size;
697 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
698 					    "==> nxge_start(12_2): "
699 					    "USE DVMA: len %d clen %d ",
700 					    len, clen));
701 
702 					i = TXDMA_DESC_NEXT_INDEX(i, 1,
703 					    tx_ring_p->tx_wrap_mask);
704 					tx_desc_p = &tx_desc_ring_vp[i];
705 
706 #if defined(__i386)
707 					npi_desc_handle.regp =
708 					    (uint32_t)tx_desc_p;
709 #else
710 					npi_desc_handle.regp =
711 					    (uint64_t)tx_desc_p;
712 #endif
713 					tx_msg_p = &tx_msg_ring[i];
714 					tx_msg_p->flags.dma_type = USE_NONE;
715 					tx_desc.value = 0;
716 
717 					ncookies--;
718 				}
719 				tdc_stats->tx_ddi_pkts++;
720 				NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start:"
721 				    "DMA: ddi packets %d",
722 				    tdc_stats->tx_ddi_pkts));
723 			} else {
724 				NXGE_ERROR_MSG((nxgep, NXGE_ERR_CTL,
725 				    "dma mapping failed for %d "
726 				    "bytes addr $%p flags %x (%d)",
727 				    len, b_rptr, status, status));
728 				good_packet = B_FALSE;
729 				tdc_stats->tx_dma_bind_fail++;
730 				tx_msg_p->flags.dma_type = USE_NONE;
731 				if (is_lso) {
732 					mp = nmp;
733 					goto nxge_start_fail_lso;
734 				} else {
735 					status = 1;
736 					goto nxge_start_fail2;
737 				}
738 			}
739 		} /* ddi dvma */
740 
741 		if (is_lso) {
742 			nmp_lso_save = nmp;
743 		}
744 		nmp = nmp->b_cont;
745 nxge_start_control_header_only:
746 #if defined(__i386)
747 		npi_desc_handle.regp = (uint32_t)tx_desc_p;
748 #else
749 		npi_desc_handle.regp = (uint64_t)tx_desc_p;
750 #endif
751 		ngathers++;
752 
753 		if (ngathers == 1) {
754 #ifdef	NXGE_DEBUG
755 			save_desc_p = &sop_tx_desc;
756 #endif
757 			sop_tx_desc_p = &sop_tx_desc;
758 			sop_tx_desc_p->value = 0;
759 			sop_tx_desc_p->bits.hdw.tr_len = clen;
760 			sop_tx_desc_p->bits.hdw.sad = dma_ioaddr >> 32;
761 			sop_tx_desc_p->bits.ldw.sad = dma_ioaddr & 0xffffffff;
762 		} else {
763 #ifdef	NXGE_DEBUG
764 			save_desc_p = &tx_desc;
765 #endif
766 			tmp_desc_p = &tx_desc;
767 			tmp_desc_p->value = 0;
768 			tmp_desc_p->bits.hdw.tr_len = clen;
769 			tmp_desc_p->bits.hdw.sad = dma_ioaddr >> 32;
770 			tmp_desc_p->bits.ldw.sad = dma_ioaddr & 0xffffffff;
771 
772 			tx_desc_p->value = tmp_desc_p->value;
773 		}
774 
775 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(13): "
776 		    "Desc_entry %d ngathers %d "
777 		    "desc_vp $%p tx_desc_p $%p "
778 		    "len %d clen %d pkt_len %d pack_len %d nmblks %d "
779 		    "dma_ioaddr (SAD) $%p mark %d",
780 		    i, ngathers,
781 		    tx_desc_ring_vp, tx_desc_p,
782 		    len, clen, pkt_len, pack_len, nmblks,
783 		    dma_ioaddr, mark_mode));
784 
785 #ifdef NXGE_DEBUG
786 		npi_desc_handle.nxgep = nxgep;
787 		npi_desc_handle.function.function = nxgep->function_num;
788 		npi_desc_handle.function.instance = nxgep->instance;
789 		sad = (save_desc_p->value & TX_PKT_DESC_SAD_MASK);
790 		xfer_len = ((save_desc_p->value & TX_PKT_DESC_TR_LEN_MASK) >>
791 		    TX_PKT_DESC_TR_LEN_SHIFT);
792 
793 
794 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "\n\t: value 0x%llx\n"
795 		    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\t"
796 		    "mark %d sop %d\n",
797 		    save_desc_p->value,
798 		    sad,
799 		    save_desc_p->bits.hdw.tr_len,
800 		    xfer_len,
801 		    save_desc_p->bits.hdw.num_ptr,
802 		    save_desc_p->bits.hdw.mark,
803 		    save_desc_p->bits.hdw.sop));
804 
805 		npi_txdma_dump_desc_one(npi_desc_handle, NULL, i);
806 #endif
807 
808 		tx_msg_p->tx_msg_size = clen;
809 		i = TXDMA_DESC_NEXT_INDEX(i, 1, tx_ring_p->tx_wrap_mask);
810 		if (ngathers > nxge_tx_max_gathers) {
811 			good_packet = B_FALSE;
812 			hcksum_retrieve(mp, NULL, NULL, &start_offset,
813 			    &stuff_offset, &end_offset, &value,
814 			    &cksum_flags);
815 
816 			NXGE_DEBUG_MSG((NULL, TX_CTL,
817 			    "==> nxge_start(14): pull msg - "
818 			    "len %d pkt_len %d ngathers %d",
819 			    len, pkt_len, ngathers));
820 
821 			/*
822 			 * Just give up on this packet.
823 			 */
824 			if (is_lso) {
825 				mp = nmp_lso_save;
826 				goto nxge_start_fail_lso;
827 			}
828 			status = 0;
829 			goto nxge_start_fail2;
830 		}
831 	} /* while (nmp) */
832 
833 	tx_msg_p->tx_message = mp;
834 	tx_desc_p = &tx_desc_ring_vp[sop_index];
835 #if defined(__i386)
836 	npi_desc_handle.regp = (uint32_t)tx_desc_p;
837 #else
838 	npi_desc_handle.regp = (uint64_t)tx_desc_p;
839 #endif
840 
841 	pkthdrp = (p_tx_pkt_hdr_all_t)hdrp;
842 	pkthdrp->reserved = 0;
843 	hdrp->value = 0;
844 	bcopy(&tmp_hdrp, hdrp, sizeof (tx_pkt_header_t));
845 
846 	if (pkt_len > NXGE_MTU_DEFAULT_MAX) {
847 		tdc_stats->tx_jumbo_pkts++;
848 	}
849 
850 	min_len = (ETHERMIN + TX_PKT_HEADER_SIZE + (npads * 2));
851 	if (pkt_len < min_len) {
852 		/* Assume we use bcopy to premapped buffers */
853 		kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
854 		NXGE_DEBUG_MSG((NULL, TX_CTL,
855 		    "==> nxge_start(14-1): < (msg_min + 16)"
856 		    "len %d pkt_len %d min_len %d bzero %d ngathers %d",
857 		    len, pkt_len, min_len, (min_len - pkt_len), ngathers));
858 		bzero((kaddr + pkt_len), (min_len - pkt_len));
859 		pkt_len = tx_msg_p->tx_msg_size = min_len;
860 
861 		sop_tx_desc_p->bits.hdw.tr_len = min_len;
862 
863 		NXGE_MEM_PIO_WRITE64(npi_desc_handle, sop_tx_desc_p->value);
864 		tx_desc_p->value = sop_tx_desc_p->value;
865 
866 		NXGE_DEBUG_MSG((NULL, TX_CTL,
867 		    "==> nxge_start(14-2): < msg_min - "
868 		    "len %d pkt_len %d min_len %d ngathers %d",
869 		    len, pkt_len, min_len, ngathers));
870 	}
871 
872 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: cksum_flags 0x%x ",
873 	    cksum_flags));
874 	{
875 		uint64_t	tmp_len;
876 
877 		/* pkt_len already includes 16 + paddings!! */
878 		/* Update the control header length */
879 		tot_xfer_len = (pkt_len - TX_PKT_HEADER_SIZE);
880 		tmp_len = hdrp->value |
881 		    (tot_xfer_len << TX_PKT_HEADER_TOT_XFER_LEN_SHIFT);
882 
883 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
884 		    "==> nxge_start(15_x1): setting SOP "
885 		    "tot_xfer_len 0x%llx (%d) pkt_len %d tmp_len "
886 		    "0x%llx hdrp->value 0x%llx",
887 		    tot_xfer_len, tot_xfer_len, pkt_len,
888 		    tmp_len, hdrp->value));
889 #if defined(_BIG_ENDIAN)
890 		hdrp->value = ddi_swap64(tmp_len);
891 #else
892 		hdrp->value = tmp_len;
893 #endif
894 		NXGE_DEBUG_MSG((nxgep,
895 		    TX_CTL, "==> nxge_start(15_x2): setting SOP "
896 		    "after SWAP: tot_xfer_len 0x%llx pkt_len %d "
897 		    "tmp_len 0x%llx hdrp->value 0x%llx",
898 		    tot_xfer_len, pkt_len,
899 		    tmp_len, hdrp->value));
900 	}
901 
902 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(15): setting SOP "
903 	    "wr_index %d "
904 	    "tot_xfer_len (%d) pkt_len %d npads %d",
905 	    sop_index,
906 	    tot_xfer_len, pkt_len,
907 	    npads));
908 
909 	sop_tx_desc_p->bits.hdw.sop = 1;
910 	sop_tx_desc_p->bits.hdw.mark = mark_mode;
911 	sop_tx_desc_p->bits.hdw.num_ptr = ngathers;
912 
913 	NXGE_MEM_PIO_WRITE64(npi_desc_handle, sop_tx_desc_p->value);
914 
915 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(16): set SOP done"));
916 
917 #ifdef NXGE_DEBUG
918 	npi_desc_handle.nxgep = nxgep;
919 	npi_desc_handle.function.function = nxgep->function_num;
920 	npi_desc_handle.function.instance = nxgep->instance;
921 
922 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "\n\t: value 0x%llx\n"
923 	    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\tmark %d sop %d\n",
924 	    save_desc_p->value,
925 	    sad,
926 	    save_desc_p->bits.hdw.tr_len,
927 	    xfer_len,
928 	    save_desc_p->bits.hdw.num_ptr,
929 	    save_desc_p->bits.hdw.mark,
930 	    save_desc_p->bits.hdw.sop));
931 	(void) npi_txdma_dump_desc_one(npi_desc_handle, NULL, sop_index);
932 
933 	dump_len = (pkt_len > 128) ? 128: pkt_len;
934 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
935 	    "==> nxge_start: dump packets(17) (after sop set, len "
936 	    " (len/dump_len/pkt_len/tot_xfer_len) %d/%d/%d/%d):\n"
937 	    "ptr $%p: %s", len, dump_len, pkt_len, tot_xfer_len,
938 	    (char *)hdrp,
939 	    nxge_dump_packet((char *)hdrp, dump_len)));
940 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
941 	    "==> nxge_start(18): TX desc sync: sop_index %d",
942 	    sop_index));
943 #endif
944 
945 	if ((ngathers == 1) || tx_ring_p->wr_index < i) {
946 		(void) ddi_dma_sync(tx_desc_dma_handle,
947 		    sop_index * sizeof (tx_desc_t),
948 		    ngathers * sizeof (tx_desc_t),
949 		    DDI_DMA_SYNC_FORDEV);
950 
951 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(19): sync 1 "
952 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
953 		    "pkt_len %d ngathers %d sop_index %d\n",
954 		    stuff_offset, start_offset,
955 		    pkt_len, ngathers, sop_index));
956 	} else { /* more than one descriptor and wrap around */
957 		uint32_t nsdescs = tx_ring_p->tx_ring_size - sop_index;
958 		(void) ddi_dma_sync(tx_desc_dma_handle,
959 		    sop_index * sizeof (tx_desc_t),
960 		    nsdescs * sizeof (tx_desc_t),
961 		    DDI_DMA_SYNC_FORDEV);
962 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(20): sync 1 "
963 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
964 		    "pkt_len %d ngathers %d sop_index %d\n",
965 		    stuff_offset, start_offset,
966 		    pkt_len, ngathers, sop_index));
967 
968 		(void) ddi_dma_sync(tx_desc_dma_handle,
969 		    0,
970 		    (ngathers - nsdescs) * sizeof (tx_desc_t),
971 		    DDI_DMA_SYNC_FORDEV);
972 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(21): sync 2 "
973 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
974 		    "pkt_len %d ngathers %d sop_index %d\n",
975 		    stuff_offset, start_offset,
976 		    pkt_len, ngathers, sop_index));
977 	}
978 
979 	tail_index = tx_ring_p->wr_index;
980 	tail_wrap = tx_ring_p->wr_index_wrap;
981 
982 	tx_ring_p->wr_index = i;
983 	if (tx_ring_p->wr_index <= tail_index) {
984 		tx_ring_p->wr_index_wrap = ((tail_wrap == B_TRUE) ?
985 		    B_FALSE : B_TRUE);
986 	}
987 
988 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: TX kick: "
989 	    "channel %d wr_index %d wrap %d ngathers %d desc_pend %d",
990 	    tx_ring_p->tdc,
991 	    tx_ring_p->wr_index,
992 	    tx_ring_p->wr_index_wrap,
993 	    ngathers,
994 	    tx_ring_p->descs_pending));
995 
996 	if (is_lso) {
997 		lso_ngathers += ngathers;
998 		if (mp_chain != NULL) {
999 			mp = mp_chain;
1000 			mp_chain = mp_chain->b_next;
1001 			mp->b_next = NULL;
1002 			if (nxge_lso_kick_cnt == lso_ngathers) {
1003 				tx_ring_p->descs_pending += lso_ngathers;
1004 				{
1005 					tx_ring_kick_t		kick;
1006 
1007 					kick.value = 0;
1008 					kick.bits.ldw.wrap =
1009 					    tx_ring_p->wr_index_wrap;
1010 					kick.bits.ldw.tail =
1011 					    (uint16_t)tx_ring_p->wr_index;
1012 
1013 					/* Kick the Transmit kick register */
1014 					TXDMA_REG_WRITE64(
1015 					    NXGE_DEV_NPI_HANDLE(nxgep),
1016 					    TX_RING_KICK_REG,
1017 					    (uint8_t)tx_ring_p->tdc,
1018 					    kick.value);
1019 					tdc_stats->tx_starts++;
1020 
1021 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
1022 					    "==> nxge_start: more LSO: "
1023 					    "LSO_CNT %d",
1024 					    lso_ngathers));
1025 				}
1026 				lso_ngathers = 0;
1027 				ngathers = 0;
1028 				cur_index_lso = sop_index = tx_ring_p->wr_index;
1029 				lso_tail_wrap = tx_ring_p->wr_index_wrap;
1030 			}
1031 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
1032 			    "==> nxge_start: lso again: "
1033 			    "lso_gathers %d ngathers %d cur_index_lso %d "
1034 			    "wr_index %d sop_index %d",
1035 			    lso_ngathers, ngathers, cur_index_lso,
1036 			    tx_ring_p->wr_index, sop_index));
1037 
1038 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
1039 			    "==> nxge_start: next : count %d",
1040 			    lso_ngathers));
1041 			lso_again = B_TRUE;
1042 			goto start_again;
1043 		}
1044 		ngathers = lso_ngathers;
1045 	}
1046 
1047 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: TX KICKING: "));
1048 
1049 	{
1050 		tx_ring_kick_t		kick;
1051 
1052 		kick.value = 0;
1053 		kick.bits.ldw.wrap = tx_ring_p->wr_index_wrap;
1054 		kick.bits.ldw.tail = (uint16_t)tx_ring_p->wr_index;
1055 
1056 		/* Kick start the Transmit kick register */
1057 		TXDMA_REG_WRITE64(NXGE_DEV_NPI_HANDLE(nxgep),
1058 		    TX_RING_KICK_REG,
1059 		    (uint8_t)tx_ring_p->tdc,
1060 		    kick.value);
1061 	}
1062 
1063 	tx_ring_p->descs_pending += ngathers;
1064 	tdc_stats->tx_starts++;
1065 
1066 	if (isLDOMservice(nxgep)) {
1067 		tx_ring_p->tx_ring_busy = B_FALSE;
1068 		if (tx_ring_p->tx_ring_offline) {
1069 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
1070 			    NXGE_TX_RING_OFFLINED);
1071 		}
1072 	}
1073 
1074 	MUTEX_EXIT(&tx_ring_p->lock);
1075 
1076 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "<== nxge_start"));
1077 	return (status);
1078 
1079 nxge_start_fail_lso:
1080 	status = 0;
1081 	good_packet = B_FALSE;
1082 	if (mp != NULL)
1083 		freemsg(mp);
1084 	if (mp_chain != NULL)
1085 		freemsgchain(mp_chain);
1086 
1087 	if (!lso_again && !ngathers) {
1088 		if (isLDOMservice(nxgep)) {
1089 			tx_ring_p->tx_ring_busy = B_FALSE;
1090 			if (tx_ring_p->tx_ring_offline) {
1091 				(void) atomic_swap_32(
1092 				    &tx_ring_p->tx_ring_offline,
1093 				    NXGE_TX_RING_OFFLINED);
1094 			}
1095 		}
1096 
1097 		MUTEX_EXIT(&tx_ring_p->lock);
1098 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
1099 		    "==> nxge_start: lso exit (nothing changed)"));
1100 		goto nxge_start_fail1;
1101 	}
1102 
1103 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
1104 	    "==> nxge_start (channel %d): before lso "
1105 	    "lso_gathers %d ngathers %d cur_index_lso %d "
1106 	    "wr_index %d sop_index %d lso_again %d",
1107 	    tx_ring_p->tdc,
1108 	    lso_ngathers, ngathers, cur_index_lso,
1109 	    tx_ring_p->wr_index, sop_index, lso_again));
1110 
1111 	if (lso_again) {
1112 		lso_ngathers += ngathers;
1113 		ngathers = lso_ngathers;
1114 		sop_index = cur_index_lso;
1115 		tx_ring_p->wr_index = sop_index;
1116 		tx_ring_p->wr_index_wrap = lso_tail_wrap;
1117 	}
1118 
1119 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
1120 	    "==> nxge_start (channel %d): after lso "
1121 	    "lso_gathers %d ngathers %d cur_index_lso %d "
1122 	    "wr_index %d sop_index %d lso_again %d",
1123 	    tx_ring_p->tdc,
1124 	    lso_ngathers, ngathers, cur_index_lso,
1125 	    tx_ring_p->wr_index, sop_index, lso_again));
1126 
1127 nxge_start_fail2:
1128 	if (good_packet == B_FALSE) {
1129 		cur_index = sop_index;
1130 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: clean up"));
1131 		for (i = 0; i < ngathers; i++) {
1132 			tx_desc_p = &tx_desc_ring_vp[cur_index];
1133 #if defined(__i386)
1134 			npi_handle.regp = (uint32_t)tx_desc_p;
1135 #else
1136 			npi_handle.regp = (uint64_t)tx_desc_p;
1137 #endif
1138 			tx_msg_p = &tx_msg_ring[cur_index];
1139 			(void) npi_txdma_desc_set_zero(npi_handle, 1);
1140 			if (tx_msg_p->flags.dma_type == USE_DVMA) {
1141 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
1142 				    "tx_desc_p = %X index = %d",
1143 				    tx_desc_p, tx_ring_p->rd_index));
1144 				(void) dvma_unload(tx_msg_p->dvma_handle,
1145 				    0, -1);
1146 				tx_msg_p->dvma_handle = NULL;
1147 				if (tx_ring_p->dvma_wr_index ==
1148 				    tx_ring_p->dvma_wrap_mask)
1149 					tx_ring_p->dvma_wr_index = 0;
1150 				else
1151 					tx_ring_p->dvma_wr_index++;
1152 				tx_ring_p->dvma_pending--;
1153 			} else if (tx_msg_p->flags.dma_type == USE_DMA) {
1154 				if (ddi_dma_unbind_handle(
1155 				    tx_msg_p->dma_handle)) {
1156 					cmn_err(CE_WARN, "!nxge_start: "
1157 					    "ddi_dma_unbind_handle failed");
1158 				}
1159 			}
1160 			tx_msg_p->flags.dma_type = USE_NONE;
1161 			cur_index = TXDMA_DESC_NEXT_INDEX(cur_index, 1,
1162 			    tx_ring_p->tx_wrap_mask);
1163 
1164 		}
1165 	}
1166 
1167 	if (isLDOMservice(nxgep)) {
1168 		tx_ring_p->tx_ring_busy = B_FALSE;
1169 		if (tx_ring_p->tx_ring_offline) {
1170 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
1171 			    NXGE_TX_RING_OFFLINED);
1172 		}
1173 	}
1174 
1175 	MUTEX_EXIT(&tx_ring_p->lock);
1176 
1177 nxge_start_fail1:
1178 	/* Add FMA to check the access handle nxge_hregh */
1179 
1180 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "<== nxge_start"));
1181 	return (status);
1182 }
1183 
1184 /* Software LSO starts here */
1185 static void
1186 nxge_hcksum_retrieve(mblk_t *mp,
1187     uint32_t *start, uint32_t *stuff, uint32_t *end,
1188     uint32_t *value, uint32_t *flags)
1189 {
1190 	if (mp->b_datap->db_type == M_DATA) {
1191 		if (flags != NULL) {
1192 			*flags = DB_CKSUMFLAGS(mp) & (HCK_IPV4_HDRCKSUM |
1193 			    HCK_PARTIALCKSUM | HCK_FULLCKSUM |
1194 			    HCK_FULLCKSUM_OK);
1195 			if ((*flags & (HCK_PARTIALCKSUM |
1196 			    HCK_FULLCKSUM)) != 0) {
1197 				if (value != NULL)
1198 					*value = (uint32_t)DB_CKSUM16(mp);
1199 				if ((*flags & HCK_PARTIALCKSUM) != 0) {
1200 					if (start != NULL)
1201 						*start =
1202 						    (uint32_t)DB_CKSUMSTART(mp);
1203 					if (stuff != NULL)
1204 						*stuff =
1205 						    (uint32_t)DB_CKSUMSTUFF(mp);
1206 					if (end != NULL)
1207 						*end =
1208 						    (uint32_t)DB_CKSUMEND(mp);
1209 				}
1210 			}
1211 		}
1212 	}
1213 }
1214 
1215 static void
1216 nxge_lso_info_get(mblk_t *mp, uint32_t *mss, uint32_t *flags)
1217 {
1218 	ASSERT(DB_TYPE(mp) == M_DATA);
1219 
1220 	*mss = 0;
1221 	if (flags != NULL) {
1222 		*flags = DB_CKSUMFLAGS(mp) & HW_LSO;
1223 		if ((*flags != 0) && (mss != NULL)) {
1224 			*mss = (uint32_t)DB_LSOMSS(mp);
1225 		}
1226 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1227 		    "==> nxge_lso_info_get(flag !=NULL): mss %d *flags 0x%x",
1228 		    *mss, *flags));
1229 	}
1230 
1231 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1232 	    "<== nxge_lso_info_get: mss %d", *mss));
1233 }
1234 
1235 /*
1236  * Do Soft LSO on the oversized packet.
1237  *
1238  * 1. Create a chain of message for headers.
1239  * 2. Fill up header messages with proper information.
1240  * 3. Copy Eithernet, IP, and TCP headers from the original message to
1241  *    each new message with necessary adjustments.
1242  *    * Unchange the ethernet header for DIX frames. (by default)
1243  *    * IP Total Length field is updated to MSS or less(only for the last one).
1244  *    * IP Identification value is incremented by one for each packet.
1245  *    * TCP sequence Number is recalculated according to the payload length.
1246  *    * Set FIN and/or PSH flags for the *last* packet if applied.
1247  *    * TCP partial Checksum
1248  * 4. Update LSO information in the first message header.
1249  * 5. Release the original message header.
1250  */
1251 static mblk_t *
1252 nxge_do_softlso(mblk_t *mp, uint32_t mss)
1253 {
1254 	uint32_t	hckflags;
1255 	int		pktlen;
1256 	int		hdrlen;
1257 	int		segnum;
1258 	int		i;
1259 	struct ether_vlan_header *evh;
1260 	int		ehlen, iphlen, tcphlen;
1261 	struct ip	*oiph, *niph;
1262 	struct tcphdr *otcph, *ntcph;
1263 	int		available, len, left;
1264 	uint16_t	ip_id;
1265 	uint32_t	tcp_seq;
1266 #ifdef __sparc
1267 	uint32_t	tcp_seq_tmp;
1268 #endif
1269 	mblk_t		*datamp;
1270 	uchar_t		*rptr;
1271 	mblk_t		*nmp;
1272 	mblk_t		*cmp;
1273 	mblk_t		*mp_chain;
1274 	boolean_t do_cleanup = B_FALSE;
1275 	t_uscalar_t start_offset = 0;
1276 	t_uscalar_t stuff_offset = 0;
1277 	t_uscalar_t value = 0;
1278 	uint16_t	l4_len;
1279 	ipaddr_t	src, dst;
1280 	uint32_t	cksum, sum, l4cksum;
1281 
1282 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1283 	    "==> nxge_do_softlso"));
1284 	/*
1285 	 * check the length of LSO packet payload and calculate the number of
1286 	 * segments to be generated.
1287 	 */
1288 	pktlen = msgsize(mp);
1289 	evh = (struct ether_vlan_header *)mp->b_rptr;
1290 
1291 	/* VLAN? */
1292 	if (evh->ether_tpid == htons(ETHERTYPE_VLAN))
1293 		ehlen = sizeof (struct ether_vlan_header);
1294 	else
1295 		ehlen = sizeof (struct ether_header);
1296 	oiph = (struct ip *)(mp->b_rptr + ehlen);
1297 	iphlen = oiph->ip_hl * 4;
1298 	otcph = (struct tcphdr *)(mp->b_rptr + ehlen + iphlen);
1299 	tcphlen = otcph->th_off * 4;
1300 
1301 	l4_len = pktlen - ehlen - iphlen;
1302 
1303 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1304 	    "==> nxge_do_softlso: mss %d oiph $%p "
1305 	    "original ip_sum oiph->ip_sum 0x%x "
1306 	    "original tcp_sum otcph->th_sum 0x%x "
1307 	    "oiph->ip_len %d pktlen %d ehlen %d "
1308 	    "l4_len %d (0x%x) ip_len - iphlen %d ",
1309 	    mss,
1310 	    oiph,
1311 	    oiph->ip_sum,
1312 	    otcph->th_sum,
1313 	    ntohs(oiph->ip_len), pktlen,
1314 	    ehlen,
1315 	    l4_len,
1316 	    l4_len,
1317 	    ntohs(oiph->ip_len) - iphlen));
1318 
1319 	/* IPv4 + TCP */
1320 	if (!(oiph->ip_v == IPV4_VERSION)) {
1321 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1322 		    "<== nxge_do_softlso: not IPV4 "
1323 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1324 		    ntohs(oiph->ip_len), pktlen, ehlen,
1325 		    tcphlen));
1326 		freemsg(mp);
1327 		return (NULL);
1328 	}
1329 
1330 	if (!(oiph->ip_p == IPPROTO_TCP)) {
1331 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1332 		    "<== nxge_do_softlso: not TCP "
1333 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1334 		    ntohs(oiph->ip_len), pktlen, ehlen,
1335 		    tcphlen));
1336 		freemsg(mp);
1337 		return (NULL);
1338 	}
1339 
1340 	if (!(ntohs(oiph->ip_len) == pktlen - ehlen)) {
1341 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1342 		    "<== nxge_do_softlso: len not matched  "
1343 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1344 		    ntohs(oiph->ip_len), pktlen, ehlen,
1345 		    tcphlen));
1346 		freemsg(mp);
1347 		return (NULL);
1348 	}
1349 
1350 	otcph = (struct tcphdr *)(mp->b_rptr + ehlen + iphlen);
1351 	tcphlen = otcph->th_off * 4;
1352 
1353 	/* TCP flags can not include URG, RST, or SYN */
1354 	VERIFY((otcph->th_flags & (TH_SYN | TH_RST | TH_URG)) == 0);
1355 
1356 	hdrlen = ehlen + iphlen + tcphlen;
1357 
1358 	VERIFY(MBLKL(mp) >= hdrlen);
1359 
1360 	if (MBLKL(mp) > hdrlen) {
1361 		datamp = mp;
1362 		rptr = mp->b_rptr + hdrlen;
1363 	} else { /* = */
1364 		datamp = mp->b_cont;
1365 		rptr = datamp->b_rptr;
1366 	}
1367 
1368 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1369 	    "nxge_do_softlso: otcph $%p pktlen: %d, "
1370 	    "hdrlen %d ehlen %d iphlen %d tcphlen %d "
1371 	    "mblkl(mp): %d, mblkl(datamp): %d",
1372 	    otcph,
1373 	    pktlen, hdrlen, ehlen, iphlen, tcphlen,
1374 	    (int)MBLKL(mp), (int)MBLKL(datamp)));
1375 
1376 	hckflags = 0;
1377 	nxge_hcksum_retrieve(mp,
1378 	    &start_offset, &stuff_offset, &value, NULL, &hckflags);
1379 
1380 	dst = oiph->ip_dst.s_addr;
1381 	src = oiph->ip_src.s_addr;
1382 
1383 	cksum = (dst >> 16) + (dst & 0xFFFF) +
1384 	    (src >> 16) + (src & 0xFFFF);
1385 	l4cksum = cksum + IP_TCP_CSUM_COMP;
1386 
1387 	sum = l4_len + l4cksum;
1388 	sum = (sum & 0xFFFF) + (sum >> 16);
1389 
1390 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1391 	    "==> nxge_do_softlso: dst 0x%x src 0x%x sum 0x%x ~new 0x%x "
1392 	    "hckflags 0x%x start_offset %d stuff_offset %d "
1393 	    "value (original) 0x%x th_sum 0x%x "
1394 	    "pktlen %d l4_len %d (0x%x) "
1395 	    "MBLKL(mp): %d, MBLKL(datamp): %d dump header %s",
1396 	    dst, src,
1397 	    (sum & 0xffff), (~sum & 0xffff),
1398 	    hckflags, start_offset, stuff_offset,
1399 	    value, otcph->th_sum,
1400 	    pktlen,
1401 	    l4_len,
1402 	    l4_len,
1403 	    ntohs(oiph->ip_len) - (int)MBLKL(mp),
1404 	    (int)MBLKL(datamp),
1405 	    nxge_dump_packet((char *)evh, 12)));
1406 
1407 	/*
1408 	 * Start to process.
1409 	 */
1410 	available = pktlen - hdrlen;
1411 	segnum = (available - 1) / mss + 1;
1412 
1413 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1414 	    "==> nxge_do_softlso: pktlen %d "
1415 	    "MBLKL(mp): %d, MBLKL(datamp): %d "
1416 	    "available %d mss %d segnum %d",
1417 	    pktlen, (int)MBLKL(mp), (int)MBLKL(datamp),
1418 	    available,
1419 	    mss,
1420 	    segnum));
1421 
1422 	VERIFY(segnum >= 2);
1423 
1424 	/*
1425 	 * Try to pre-allocate all header messages
1426 	 */
1427 	mp_chain = NULL;
1428 	for (i = 0; i < segnum; i++) {
1429 		if ((nmp = allocb(hdrlen, 0)) == NULL) {
1430 			/* Clean up the mp_chain */
1431 			while (mp_chain != NULL) {
1432 				nmp = mp_chain;
1433 				mp_chain = mp_chain->b_next;
1434 				freemsg(nmp);
1435 			}
1436 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1437 			    "<== nxge_do_softlso: "
1438 			    "Could not allocate enough messages for headers!"));
1439 			freemsg(mp);
1440 			return (NULL);
1441 		}
1442 		nmp->b_next = mp_chain;
1443 		mp_chain = nmp;
1444 
1445 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1446 		    "==> nxge_do_softlso: "
1447 		    "mp $%p nmp $%p mp_chain $%p mp_chain->b_next $%p",
1448 		    mp, nmp, mp_chain, mp_chain->b_next));
1449 	}
1450 
1451 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1452 	    "==> nxge_do_softlso: mp $%p nmp $%p mp_chain $%p",
1453 	    mp, nmp, mp_chain));
1454 
1455 	/*
1456 	 * Associate payload with new packets
1457 	 */
1458 	cmp = mp_chain;
1459 	left = available;
1460 	while (cmp != NULL) {
1461 		nmp = dupb(datamp);
1462 		if (nmp == NULL) {
1463 			do_cleanup = B_TRUE;
1464 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1465 			    "==>nxge_do_softlso: "
1466 			    "Can not dupb(datamp), have to do clean up"));
1467 			goto cleanup_allocated_msgs;
1468 		}
1469 
1470 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1471 		    "==> nxge_do_softlso: (loop) before mp $%p cmp $%p "
1472 		    "dupb nmp $%p len %d left %d msd %d ",
1473 		    mp, cmp, nmp, len, left, mss));
1474 
1475 		cmp->b_cont = nmp;
1476 		nmp->b_rptr = rptr;
1477 		len = (left < mss) ? left : mss;
1478 		left -= len;
1479 
1480 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1481 		    "==> nxge_do_softlso: (loop) after mp $%p cmp $%p "
1482 		    "dupb nmp $%p len %d left %d mss %d ",
1483 		    mp, cmp, nmp, len, left, mss));
1484 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1485 		    "nxge_do_softlso: before available: %d, "
1486 		    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1487 		    available, left, len, segnum, (int)MBLKL(nmp)));
1488 
1489 		len -= MBLKL(nmp);
1490 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1491 		    "nxge_do_softlso: after available: %d, "
1492 		    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1493 		    available, left, len, segnum, (int)MBLKL(nmp)));
1494 
1495 		while (len > 0) {
1496 			mblk_t *mmp = NULL;
1497 
1498 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1499 			    "nxge_do_softlso: (4) len > 0 available: %d, "
1500 			    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1501 			    available, left, len, segnum, (int)MBLKL(nmp)));
1502 
1503 			if (datamp->b_cont != NULL) {
1504 				datamp = datamp->b_cont;
1505 				rptr = datamp->b_rptr;
1506 				mmp = dupb(datamp);
1507 				if (mmp == NULL) {
1508 					do_cleanup = B_TRUE;
1509 					NXGE_DEBUG_MSG((NULL, TX_CTL,
1510 					    "==> nxge_do_softlso: "
1511 					    "Can not dupb(datamp) (1), :"
1512 					    "have to do clean up"));
1513 					NXGE_DEBUG_MSG((NULL, TX_CTL,
1514 					    "==> nxge_do_softlso: "
1515 					    "available: %d, left: %d, "
1516 					    "len: %d, MBLKL(nmp): %d",
1517 					    available, left, len,
1518 					    (int)MBLKL(nmp)));
1519 					goto cleanup_allocated_msgs;
1520 				}
1521 			} else {
1522 				NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1523 				    "==> nxge_do_softlso: "
1524 				    "(1)available: %d, left: %d, "
1525 				    "len: %d, MBLKL(nmp): %d",
1526 				    available, left, len,
1527 				    (int)MBLKL(nmp)));
1528 				cmn_err(CE_PANIC,
1529 				    "==> nxge_do_softlso: "
1530 				    "Pointers must have been corrupted!\n"
1531 				    "datamp: $%p, nmp: $%p, rptr: $%p",
1532 				    (void *)datamp,
1533 				    (void *)nmp,
1534 				    (void *)rptr);
1535 			}
1536 			nmp->b_cont = mmp;
1537 			nmp = mmp;
1538 			len -= MBLKL(nmp);
1539 		}
1540 		if (len < 0) {
1541 			nmp->b_wptr += len;
1542 			rptr = nmp->b_wptr;
1543 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1544 			    "(5) len < 0 (less than 0)"
1545 			    "available: %d, left: %d, len: %d, MBLKL(nmp): %d",
1546 			    available, left, len, (int)MBLKL(nmp)));
1547 
1548 		} else if (len == 0) {
1549 			if (datamp->b_cont != NULL) {
1550 				NXGE_DEBUG_MSG((NULL, TX_CTL,
1551 				    "(5) len == 0"
1552 				    "available: %d, left: %d, len: %d, "
1553 				    "MBLKL(nmp): %d",
1554 				    available, left, len, (int)MBLKL(nmp)));
1555 				datamp = datamp->b_cont;
1556 				rptr = datamp->b_rptr;
1557 			} else {
1558 				NXGE_DEBUG_MSG((NULL, TX_CTL,
1559 				    "(6)available b_cont == NULL : %d, "
1560 				    "left: %d, len: %d, MBLKL(nmp): %d",
1561 				    available, left, len, (int)MBLKL(nmp)));
1562 
1563 				VERIFY(cmp->b_next == NULL);
1564 				VERIFY(left == 0);
1565 				break; /* Done! */
1566 			}
1567 		}
1568 		cmp = cmp->b_next;
1569 
1570 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1571 		    "(7) do_softlso: "
1572 		    "next mp in mp_chain available len != 0 : %d, "
1573 		    "left: %d, len: %d, MBLKL(nmp): %d",
1574 		    available, left, len, (int)MBLKL(nmp)));
1575 	}
1576 
1577 	/*
1578 	 * From now, start to fill up all headers for the first message
1579 	 * Hardware checksum flags need to be updated separately for FULLCKSUM
1580 	 * and PARTIALCKSUM cases. For full checksum, copy the original flags
1581 	 * into every new packet is enough. But for HCK_PARTIALCKSUM, all
1582 	 * required fields need to be updated properly.
1583 	 */
1584 	nmp = mp_chain;
1585 	bcopy(mp->b_rptr, nmp->b_rptr, hdrlen);
1586 	nmp->b_wptr = nmp->b_rptr + hdrlen;
1587 	niph = (struct ip *)(nmp->b_rptr + ehlen);
1588 	niph->ip_len = htons(mss + iphlen + tcphlen);
1589 	ip_id = ntohs(niph->ip_id);
1590 	ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1591 #ifdef __sparc
1592 	bcopy((char *)&ntcph->th_seq, &tcp_seq_tmp, 4);
1593 	tcp_seq = ntohl(tcp_seq_tmp);
1594 #else
1595 	tcp_seq = ntohl(ntcph->th_seq);
1596 #endif
1597 
1598 	ntcph->th_flags &= ~(TH_FIN | TH_PUSH | TH_RST);
1599 
1600 	DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1601 	DB_CKSUMSTART(nmp) = start_offset;
1602 	DB_CKSUMSTUFF(nmp) = stuff_offset;
1603 
1604 	/* calculate IP checksum and TCP pseudo header checksum */
1605 	niph->ip_sum = 0;
1606 	niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1607 
1608 	l4_len = mss + tcphlen;
1609 	sum = htons(l4_len) + l4cksum;
1610 	sum = (sum & 0xFFFF) + (sum >> 16);
1611 	ntcph->th_sum = (sum & 0xffff);
1612 
1613 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1614 	    "==> nxge_do_softlso: first mp $%p (mp_chain $%p) "
1615 	    "mss %d pktlen %d l4_len %d (0x%x) "
1616 	    "MBLKL(mp): %d, MBLKL(datamp): %d "
1617 	    "ip_sum 0x%x "
1618 	    "th_sum 0x%x sum 0x%x ) "
1619 	    "dump first ip->tcp %s",
1620 	    nmp, mp_chain,
1621 	    mss,
1622 	    pktlen,
1623 	    l4_len,
1624 	    l4_len,
1625 	    (int)MBLKL(mp), (int)MBLKL(datamp),
1626 	    niph->ip_sum,
1627 	    ntcph->th_sum,
1628 	    sum,
1629 	    nxge_dump_packet((char *)niph, 52)));
1630 
1631 	cmp = nmp;
1632 	while ((nmp = nmp->b_next)->b_next != NULL) {
1633 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1634 		    "==>nxge_do_softlso: middle l4_len %d ", l4_len));
1635 		bcopy(cmp->b_rptr, nmp->b_rptr, hdrlen);
1636 		nmp->b_wptr = nmp->b_rptr + hdrlen;
1637 		niph = (struct ip *)(nmp->b_rptr + ehlen);
1638 		niph->ip_id = htons(++ip_id);
1639 		niph->ip_len = htons(mss + iphlen + tcphlen);
1640 		ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1641 		tcp_seq += mss;
1642 
1643 		ntcph->th_flags &= ~(TH_FIN | TH_PUSH | TH_RST | TH_URG);
1644 
1645 #ifdef __sparc
1646 		tcp_seq_tmp = htonl(tcp_seq);
1647 		bcopy(&tcp_seq_tmp, (char *)&ntcph->th_seq, 4);
1648 #else
1649 		ntcph->th_seq = htonl(tcp_seq);
1650 #endif
1651 		DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1652 		DB_CKSUMSTART(nmp) = start_offset;
1653 		DB_CKSUMSTUFF(nmp) = stuff_offset;
1654 
1655 		/* calculate IP checksum and TCP pseudo header checksum */
1656 		niph->ip_sum = 0;
1657 		niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1658 		ntcph->th_sum = (sum & 0xffff);
1659 
1660 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1661 		    "==> nxge_do_softlso: middle ip_sum 0x%x "
1662 		    "th_sum 0x%x "
1663 		    " mp $%p (mp_chain $%p) pktlen %d "
1664 		    "MBLKL(mp): %d, MBLKL(datamp): %d ",
1665 		    niph->ip_sum,
1666 		    ntcph->th_sum,
1667 		    nmp, mp_chain,
1668 		    pktlen, (int)MBLKL(mp), (int)MBLKL(datamp)));
1669 	}
1670 
1671 	/* Last segment */
1672 	/*
1673 	 * Set FIN and/or PSH flags if present only in the last packet.
1674 	 * The ip_len could be different from prior packets.
1675 	 */
1676 	bcopy(cmp->b_rptr, nmp->b_rptr, hdrlen);
1677 	nmp->b_wptr = nmp->b_rptr + hdrlen;
1678 	niph = (struct ip *)(nmp->b_rptr + ehlen);
1679 	niph->ip_id = htons(++ip_id);
1680 	niph->ip_len = htons(msgsize(nmp->b_cont) + iphlen + tcphlen);
1681 	ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1682 	tcp_seq += mss;
1683 #ifdef __sparc
1684 	tcp_seq_tmp = htonl(tcp_seq);
1685 	bcopy(&tcp_seq_tmp, (char *)&ntcph->th_seq, 4);
1686 #else
1687 	ntcph->th_seq = htonl(tcp_seq);
1688 #endif
1689 	ntcph->th_flags = (otcph->th_flags & ~TH_URG);
1690 
1691 	DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1692 	DB_CKSUMSTART(nmp) = start_offset;
1693 	DB_CKSUMSTUFF(nmp) = stuff_offset;
1694 
1695 	/* calculate IP checksum and TCP pseudo header checksum */
1696 	niph->ip_sum = 0;
1697 	niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1698 
1699 	l4_len = ntohs(niph->ip_len) - iphlen;
1700 	sum = htons(l4_len) + l4cksum;
1701 	sum = (sum & 0xFFFF) + (sum >> 16);
1702 	ntcph->th_sum = (sum & 0xffff);
1703 
1704 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1705 	    "==> nxge_do_softlso: last next "
1706 	    "niph->ip_sum 0x%x "
1707 	    "ntcph->th_sum 0x%x sum 0x%x "
1708 	    "dump last ip->tcp %s "
1709 	    "cmp $%p mp $%p (mp_chain $%p) pktlen %d (0x%x) "
1710 	    "l4_len %d (0x%x) "
1711 	    "MBLKL(mp): %d, MBLKL(datamp): %d ",
1712 	    niph->ip_sum,
1713 	    ntcph->th_sum, sum,
1714 	    nxge_dump_packet((char *)niph, 52),
1715 	    cmp, nmp, mp_chain,
1716 	    pktlen, pktlen,
1717 	    l4_len,
1718 	    l4_len,
1719 	    (int)MBLKL(mp), (int)MBLKL(datamp)));
1720 
1721 cleanup_allocated_msgs:
1722 	if (do_cleanup) {
1723 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1724 		    "==> nxge_do_softlso: "
1725 		    "Failed allocating messages, "
1726 		    "have to clean up and fail!"));
1727 		while (mp_chain != NULL) {
1728 			nmp = mp_chain;
1729 			mp_chain = mp_chain->b_next;
1730 			freemsg(nmp);
1731 		}
1732 	}
1733 	/*
1734 	 * We're done here, so just free the original message and return the
1735 	 * new message chain, that could be NULL if failed, back to the caller.
1736 	 */
1737 	freemsg(mp);
1738 
1739 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1740 	    "<== nxge_do_softlso:mp_chain $%p", mp_chain));
1741 	return (mp_chain);
1742 }
1743 
1744 /*
1745  * Will be called before NIC driver do further operation on the message.
1746  * The input message may include LSO information, if so, go to softlso logic
1747  * to eliminate the oversized LSO packet for the incapable underlying h/w.
1748  * The return could be the same non-LSO message or a message chain for LSO case.
1749  *
1750  * The driver needs to call this function per packet and process the whole chain
1751  * if applied.
1752  */
1753 static mblk_t *
1754 nxge_lso_eliminate(mblk_t *mp)
1755 {
1756 	uint32_t lsoflags;
1757 	uint32_t mss;
1758 
1759 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1760 	    "==>nxge_lso_eliminate:"));
1761 	nxge_lso_info_get(mp, &mss, &lsoflags);
1762 
1763 	if (lsoflags & HW_LSO) {
1764 		mblk_t *nmp;
1765 
1766 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1767 		    "==>nxge_lso_eliminate:"
1768 		    "HW_LSO:mss %d mp $%p",
1769 		    mss, mp));
1770 		if ((nmp = nxge_do_softlso(mp, mss)) != NULL) {
1771 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1772 			    "<== nxge_lso_eliminate: "
1773 			    "LSO: nmp not NULL nmp $%p mss %d mp $%p",
1774 			    nmp, mss, mp));
1775 			return (nmp);
1776 		} else {
1777 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1778 			    "<== nxge_lso_eliminate_ "
1779 			    "LSO: failed nmp NULL nmp $%p mss %d mp $%p",
1780 			    nmp, mss, mp));
1781 			return (NULL);
1782 		}
1783 	}
1784 
1785 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1786 	    "<== nxge_lso_eliminate"));
1787 	return (mp);
1788 }
1789 
1790 static uint32_t
1791 nxge_csgen(uint16_t *adr, int len)
1792 {
1793 	int		i, odd;
1794 	uint32_t	sum = 0;
1795 	uint32_t	c = 0;
1796 
1797 	odd = len % 2;
1798 	for (i = 0; i < (len / 2); i++) {
1799 		sum += (adr[i] & 0xffff);
1800 	}
1801 	if (odd) {
1802 		sum += adr[len / 2] & 0xff00;
1803 	}
1804 	while ((c = ((sum & 0xffff0000) >> 16)) != 0) {
1805 		sum &= 0xffff;
1806 		sum += c;
1807 	}
1808 	return (~sum & 0xffff);
1809 }
1810