xref: /titanic_44/usr/src/uts/common/io/nxge/nxge_send.c (revision 1e4c938b57d1656808e4112127ff1dce3eba5314)
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 2010 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 	ASSERT(ring->tx_ring_handle != NULL);
66 
67 	MUTEX_ENTER(&ring->lock);
68 	(void) nxge_txdma_reclaim(ring->nxgep, ring, 0);
69 	MUTEX_EXIT(&ring->lock);
70 
71 	if (!ring->tx_ring_offline) {
72 		mac_tx_ring_update(ring->nxgep->mach, ring->tx_ring_handle);
73 	}
74 }
75 
76 static void
77 nxge_tx_ring_dispatch(p_tx_ring_t ring)
78 {
79 	/*
80 	 * Kick the ring task to reclaim some buffers.
81 	 */
82 	(void) ddi_taskq_dispatch(ring->taskq,
83 	    nxge_tx_ring_task, (void *)ring, DDI_SLEEP);
84 }
85 
86 mblk_t *
87 nxge_tx_ring_send(void *arg, mblk_t *mp)
88 {
89 	p_nxge_ring_handle_t	nrhp = (p_nxge_ring_handle_t)arg;
90 	p_nxge_t		nxgep;
91 	p_tx_ring_t		tx_ring_p;
92 	int			status, channel;
93 
94 	ASSERT(nrhp != NULL);
95 	nxgep = nrhp->nxgep;
96 	channel = nxgep->pt_config.hw_config.tdc.start + nrhp->index;
97 	tx_ring_p = nxgep->tx_rings->rings[channel];
98 
99 	/*
100 	 * We may be in a transition from offlined DMA to onlined
101 	 * DMA.
102 	 */
103 	if (tx_ring_p == NULL) {
104 		ASSERT(tx_ring_p != NULL);
105 		freemsg(mp);
106 		return ((mblk_t *)NULL);
107 	}
108 
109 	/*
110 	 * Valid DMA?
111 	 */
112 	ASSERT(nxgep == tx_ring_p->nxgep);
113 
114 	/*
115 	 * Make sure DMA is not offlined.
116 	 */
117 	if (isLDOMservice(nxgep) && tx_ring_p->tx_ring_offline) {
118 		ASSERT(!tx_ring_p->tx_ring_offline);
119 		freemsg(mp);
120 		return ((mblk_t *)NULL);
121 	}
122 
123 	/*
124 	 * Transmit the packet.
125 	 */
126 	status = nxge_start(nxgep, tx_ring_p, mp);
127 	if (status) {
128 		nxge_tx_ring_dispatch(tx_ring_p);
129 		return (mp);
130 	}
131 
132 	return ((mblk_t *)NULL);
133 }
134 
135 int
136 nxge_start(p_nxge_t nxgep, p_tx_ring_t tx_ring_p, p_mblk_t mp)
137 {
138 	int 			dma_status, status = 0;
139 	p_tx_desc_t 		tx_desc_ring_vp;
140 	npi_handle_t		npi_desc_handle;
141 	nxge_os_dma_handle_t 	tx_desc_dma_handle;
142 	p_tx_desc_t 		tx_desc_p;
143 	p_tx_msg_t 		tx_msg_ring;
144 	p_tx_msg_t 		tx_msg_p;
145 	tx_desc_t		tx_desc, *tmp_desc_p;
146 	tx_desc_t		sop_tx_desc, *sop_tx_desc_p;
147 	p_tx_pkt_header_t	hdrp;
148 	tx_pkt_hdr_all_t	tmp_hdrp;
149 	p_tx_pkt_hdr_all_t	pkthdrp;
150 	uint8_t			npads = 0;
151 	uint64_t 		dma_ioaddr;
152 	uint32_t		dma_flags;
153 	int			last_bidx;
154 	uint8_t 		*b_rptr;
155 	caddr_t 		kaddr;
156 	uint32_t		nmblks;
157 	uint32_t		ngathers;
158 	uint32_t		clen;
159 	int 			len;
160 	uint32_t		pkt_len, pack_len, min_len;
161 	uint32_t		bcopy_thresh;
162 	int 			i, cur_index, sop_index;
163 	uint16_t		tail_index;
164 	boolean_t		tail_wrap = B_FALSE;
165 	nxge_dma_common_t	desc_area;
166 	nxge_os_dma_handle_t 	dma_handle;
167 	ddi_dma_cookie_t 	dma_cookie;
168 	npi_handle_t		npi_handle;
169 	p_mblk_t 		nmp;
170 	p_mblk_t		t_mp;
171 	uint32_t 		ncookies;
172 	boolean_t 		good_packet;
173 	boolean_t 		mark_mode = B_FALSE;
174 	p_nxge_stats_t 		statsp;
175 	p_nxge_tx_ring_stats_t tdc_stats;
176 	t_uscalar_t 		start_offset = 0;
177 	t_uscalar_t 		stuff_offset = 0;
178 	t_uscalar_t 		end_offset = 0;
179 	t_uscalar_t 		value = 0;
180 	t_uscalar_t 		cksum_flags = 0;
181 	boolean_t		cksum_on = B_FALSE;
182 	uint32_t		boff = 0;
183 	uint64_t		tot_xfer_len = 0;
184 	boolean_t		header_set = B_FALSE;
185 #ifdef NXGE_DEBUG
186 	p_tx_desc_t 		tx_desc_ring_pp;
187 	p_tx_desc_t 		tx_desc_pp;
188 	tx_desc_t		*save_desc_p;
189 	int			dump_len;
190 	int			sad_len;
191 	uint64_t		sad;
192 	int			xfer_len;
193 	uint32_t		msgsize;
194 #endif
195 	p_mblk_t 		mp_chain = NULL;
196 	boolean_t		is_lso = B_FALSE;
197 	boolean_t		lso_again;
198 	int			cur_index_lso;
199 	p_mblk_t 		nmp_lso_save;
200 	uint32_t		lso_ngathers;
201 	boolean_t		lso_tail_wrap = B_FALSE;
202 
203 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
204 	    "==> nxge_start: tx dma channel %d", tx_ring_p->tdc));
205 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
206 	    "==> nxge_start: Starting tdc %d desc pending %d",
207 	    tx_ring_p->tdc, tx_ring_p->descs_pending));
208 
209 	statsp = nxgep->statsp;
210 
211 	if (!isLDOMguest(nxgep)) {
212 		switch (nxgep->mac.portmode) {
213 		default:
214 			if (nxgep->statsp->port_stats.lb_mode ==
215 			    nxge_lb_normal) {
216 				if (!statsp->mac_stats.link_up) {
217 					freemsg(mp);
218 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
219 					    "==> nxge_start: "
220 					    "link not up"));
221 					goto nxge_start_fail1;
222 				}
223 			}
224 			break;
225 		case PORT_10G_FIBER:
226 			/*
227 			 * For the following modes, check the link status
228 			 * before sending the packet out:
229 			 * nxge_lb_normal,
230 			 * nxge_lb_ext10g,
231 			 * nxge_lb_ext1000,
232 			 * nxge_lb_ext100,
233 			 * nxge_lb_ext10.
234 			 */
235 			if (nxgep->statsp->port_stats.lb_mode <
236 			    nxge_lb_phy10g) {
237 				if (!statsp->mac_stats.link_up) {
238 					freemsg(mp);
239 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
240 					    "==> nxge_start: "
241 					    "link not up"));
242 					goto nxge_start_fail1;
243 				}
244 			}
245 			break;
246 		}
247 	}
248 
249 	if ((!(nxgep->drv_state & STATE_HW_INITIALIZED)) ||
250 	    (nxgep->nxge_mac_state != NXGE_MAC_STARTED)) {
251 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
252 		    "==> nxge_start: hardware not initialized or stopped"));
253 		freemsg(mp);
254 		goto nxge_start_fail1;
255 	}
256 
257 	if (nxgep->soft_lso_enable) {
258 		mp_chain = nxge_lso_eliminate(mp);
259 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
260 		    "==> nxge_start(0): LSO mp $%p mp_chain $%p",
261 		    mp, mp_chain));
262 		if (mp_chain == NULL) {
263 			NXGE_ERROR_MSG((nxgep, TX_CTL,
264 			    "==> nxge_send(0): NULL mp_chain $%p != mp $%p",
265 			    mp_chain, mp));
266 			goto nxge_start_fail1;
267 		}
268 		if (mp_chain != mp) {
269 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
270 			    "==> nxge_send(1): IS LSO mp_chain $%p != mp $%p",
271 			    mp_chain, mp));
272 			is_lso = B_TRUE;
273 			mp = mp_chain;
274 			mp_chain = mp_chain->b_next;
275 			mp->b_next = NULL;
276 		}
277 	}
278 
279 	mac_hcksum_get(mp, &start_offset, &stuff_offset, &end_offset,
280 	    &value, &cksum_flags);
281 	if (!NXGE_IS_VLAN_PACKET(mp->b_rptr)) {
282 		start_offset += sizeof (ether_header_t);
283 		stuff_offset += sizeof (ether_header_t);
284 	} else {
285 		start_offset += sizeof (struct ether_vlan_header);
286 		stuff_offset += sizeof (struct ether_vlan_header);
287 	}
288 
289 	if (cksum_flags & HCK_PARTIALCKSUM) {
290 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
291 		    "==> nxge_start: mp $%p len %d "
292 		    "cksum_flags 0x%x (partial checksum) ",
293 		    mp, MBLKL(mp), cksum_flags));
294 		cksum_on = B_TRUE;
295 	}
296 
297 	pkthdrp = (p_tx_pkt_hdr_all_t)&tmp_hdrp;
298 	pkthdrp->reserved = 0;
299 	tmp_hdrp.pkthdr.value = 0;
300 	nxge_fill_tx_hdr(mp, B_FALSE, cksum_on,
301 	    0, 0, pkthdrp,
302 	    start_offset, stuff_offset);
303 
304 	lso_again = B_FALSE;
305 	lso_ngathers = 0;
306 
307 	MUTEX_ENTER(&tx_ring_p->lock);
308 
309 	if (isLDOMservice(nxgep)) {
310 		tx_ring_p->tx_ring_busy = B_TRUE;
311 		if (tx_ring_p->tx_ring_offline) {
312 			freemsg(mp);
313 			tx_ring_p->tx_ring_busy = B_FALSE;
314 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
315 			    NXGE_TX_RING_OFFLINED);
316 			MUTEX_EXIT(&tx_ring_p->lock);
317 			return (status);
318 		}
319 	}
320 
321 	cur_index_lso = tx_ring_p->wr_index;
322 	lso_tail_wrap = tx_ring_p->wr_index_wrap;
323 start_again:
324 	ngathers = 0;
325 	sop_index = tx_ring_p->wr_index;
326 #ifdef	NXGE_DEBUG
327 	if (tx_ring_p->descs_pending) {
328 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
329 		    "desc pending %d ", tx_ring_p->descs_pending));
330 	}
331 
332 	dump_len = (int)(MBLKL(mp));
333 	dump_len = (dump_len > 128) ? 128: dump_len;
334 
335 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
336 	    "==> nxge_start: tdc %d: dumping ...: b_rptr $%p "
337 	    "(Before header reserve: ORIGINAL LEN %d)",
338 	    tx_ring_p->tdc,
339 	    mp->b_rptr,
340 	    dump_len));
341 
342 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: dump packets "
343 	    "(IP ORIGINAL b_rptr $%p): %s", mp->b_rptr,
344 	    nxge_dump_packet((char *)mp->b_rptr, dump_len)));
345 #endif
346 
347 	tdc_stats = tx_ring_p->tdc_stats;
348 	mark_mode = (tx_ring_p->descs_pending &&
349 	    (((int)tx_ring_p->tx_ring_size - (int)tx_ring_p->descs_pending) <
350 	    (int)nxge_tx_minfree));
351 
352 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
353 	    "TX Descriptor ring is channel %d mark mode %d",
354 	    tx_ring_p->tdc, mark_mode));
355 
356 	if ((tx_ring_p->descs_pending + lso_ngathers) >= nxge_reclaim_pending) {
357 		if (!nxge_txdma_reclaim(nxgep, tx_ring_p,
358 		    (nxge_tx_minfree + lso_ngathers))) {
359 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
360 			    "TX Descriptor ring is full: channel %d",
361 			    tx_ring_p->tdc));
362 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
363 			    "TX Descriptor ring is full: channel %d",
364 			    tx_ring_p->tdc));
365 			if (is_lso) {
366 				/*
367 				 * free the current mp and mp_chain if not FULL.
368 				 */
369 				tdc_stats->tx_no_desc++;
370 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
371 				    "LSO packet: TX Descriptor ring is full: "
372 				    "channel %d",
373 				    tx_ring_p->tdc));
374 				goto nxge_start_fail_lso;
375 			} else {
376 				(void) cas32((uint32_t *)&tx_ring_p->queueing,
377 				    0, 1);
378 				tdc_stats->tx_no_desc++;
379 
380 				if (isLDOMservice(nxgep)) {
381 					tx_ring_p->tx_ring_busy = B_FALSE;
382 					if (tx_ring_p->tx_ring_offline) {
383 						(void) atomic_swap_32(
384 						    &tx_ring_p->tx_ring_offline,
385 						    NXGE_TX_RING_OFFLINED);
386 					}
387 				}
388 
389 				MUTEX_EXIT(&tx_ring_p->lock);
390 				status = 1;
391 				goto nxge_start_fail1;
392 			}
393 		}
394 	}
395 
396 	nmp = mp;
397 	i = sop_index = tx_ring_p->wr_index;
398 	nmblks = 0;
399 	ngathers = 0;
400 	pkt_len = 0;
401 	pack_len = 0;
402 	clen = 0;
403 	last_bidx = -1;
404 	good_packet = B_TRUE;
405 
406 	desc_area = tx_ring_p->tdc_desc;
407 	npi_handle = desc_area.npi_handle;
408 	npi_desc_handle.regh = (nxge_os_acc_handle_t)
409 	    DMA_COMMON_ACC_HANDLE(desc_area);
410 	tx_desc_ring_vp = (p_tx_desc_t)DMA_COMMON_VPTR(desc_area);
411 	tx_desc_dma_handle = (nxge_os_dma_handle_t)
412 	    DMA_COMMON_HANDLE(desc_area);
413 	tx_msg_ring = tx_ring_p->tx_msg_ring;
414 
415 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: wr_index %d i %d",
416 	    sop_index, i));
417 
418 #ifdef	NXGE_DEBUG
419 	msgsize = msgdsize(nmp);
420 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
421 	    "==> nxge_start(1): wr_index %d i %d msgdsize %d",
422 	    sop_index, i, msgsize));
423 #endif
424 	/*
425 	 * The first 16 bytes of the premapped buffer are reserved
426 	 * for header. No padding will be used.
427 	 */
428 	pkt_len = pack_len = boff = TX_PKT_HEADER_SIZE;
429 	if (nxge_tx_use_bcopy && (nxgep->niu_type != N2_NIU)) {
430 		bcopy_thresh = (nxge_bcopy_thresh - TX_PKT_HEADER_SIZE);
431 	} else {
432 		bcopy_thresh = (TX_BCOPY_SIZE - TX_PKT_HEADER_SIZE);
433 	}
434 	while (nmp) {
435 		good_packet = B_TRUE;
436 		b_rptr = nmp->b_rptr;
437 		len = MBLKL(nmp);
438 		if (len <= 0) {
439 			nmp = nmp->b_cont;
440 			continue;
441 		}
442 		nmblks++;
443 
444 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(1): nmblks %d "
445 		    "len %d pkt_len %d pack_len %d",
446 		    nmblks, len, pkt_len, pack_len));
447 		/*
448 		 * Hardware limits the transfer length to 4K for NIU and
449 		 * 4076 (TX_MAX_TRANSFER_LENGTH) for Neptune. But we just
450 		 * use TX_MAX_TRANSFER_LENGTH as the limit for both.
451 		 * If len is longer than the limit, then we break nmp into
452 		 * two chunks: Make the first chunk equal to the limit and
453 		 * the second chunk for the remaining data. If the second
454 		 * chunk is still larger than the limit, then it will be
455 		 * broken into two in the next pass.
456 		 */
457 		if (len > TX_MAX_TRANSFER_LENGTH - TX_PKT_HEADER_SIZE) {
458 			if ((t_mp = dupb(nmp)) != NULL) {
459 				nmp->b_wptr = nmp->b_rptr +
460 				    (TX_MAX_TRANSFER_LENGTH
461 				    - TX_PKT_HEADER_SIZE);
462 				t_mp->b_rptr = nmp->b_wptr;
463 				t_mp->b_cont = nmp->b_cont;
464 				nmp->b_cont = t_mp;
465 				len = MBLKL(nmp);
466 			} else {
467 				if (is_lso) {
468 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
469 					    "LSO packet: dupb failed: "
470 					    "channel %d",
471 					    tx_ring_p->tdc));
472 					mp = nmp;
473 					goto nxge_start_fail_lso;
474 				} else {
475 					good_packet = B_FALSE;
476 					goto nxge_start_fail2;
477 				}
478 			}
479 		}
480 		tx_desc.value = 0;
481 		tx_desc_p = &tx_desc_ring_vp[i];
482 #ifdef	NXGE_DEBUG
483 		tx_desc_pp = &tx_desc_ring_pp[i];
484 #endif
485 		tx_msg_p = &tx_msg_ring[i];
486 #if defined(__i386)
487 		npi_desc_handle.regp = (uint32_t)tx_desc_p;
488 #else
489 		npi_desc_handle.regp = (uint64_t)tx_desc_p;
490 #endif
491 		if (!header_set &&
492 		    ((!nxge_tx_use_bcopy && (len > TX_BCOPY_SIZE)) ||
493 		    (len >= bcopy_thresh))) {
494 			header_set = B_TRUE;
495 			bcopy_thresh += TX_PKT_HEADER_SIZE;
496 			boff = 0;
497 			pack_len = 0;
498 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
499 			hdrp = (p_tx_pkt_header_t)kaddr;
500 			clen = pkt_len;
501 			dma_handle = tx_msg_p->buf_dma_handle;
502 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
503 			(void) ddi_dma_sync(dma_handle,
504 			    i * nxge_bcopy_thresh, nxge_bcopy_thresh,
505 			    DDI_DMA_SYNC_FORDEV);
506 
507 			tx_msg_p->flags.dma_type = USE_BCOPY;
508 			goto nxge_start_control_header_only;
509 		}
510 
511 		pkt_len += len;
512 		pack_len += len;
513 
514 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(3): "
515 		    "desc entry %d "
516 		    "DESC IOADDR $%p "
517 		    "desc_vp $%p tx_desc_p $%p "
518 		    "desc_pp $%p tx_desc_pp $%p "
519 		    "len %d pkt_len %d pack_len %d",
520 		    i,
521 		    DMA_COMMON_IOADDR(desc_area),
522 		    tx_desc_ring_vp, tx_desc_p,
523 		    tx_desc_ring_pp, tx_desc_pp,
524 		    len, pkt_len, pack_len));
525 
526 		if (len < bcopy_thresh) {
527 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(4): "
528 			    "USE BCOPY: "));
529 			if (nxge_tx_tiny_pack) {
530 				uint32_t blst =
531 				    TXDMA_DESC_NEXT_INDEX(i, -1,
532 				    tx_ring_p->tx_wrap_mask);
533 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
534 				    "==> nxge_start(5): pack"));
535 				if ((pack_len <= bcopy_thresh) &&
536 				    (last_bidx == blst)) {
537 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
538 					    "==> nxge_start: pack(6) "
539 					    "(pkt_len %d pack_len %d)",
540 					    pkt_len, pack_len));
541 					i = blst;
542 					tx_desc_p = &tx_desc_ring_vp[i];
543 #ifdef	NXGE_DEBUG
544 					tx_desc_pp = &tx_desc_ring_pp[i];
545 #endif
546 					tx_msg_p = &tx_msg_ring[i];
547 					boff = pack_len - len;
548 					ngathers--;
549 				} else if (pack_len > bcopy_thresh &&
550 				    header_set) {
551 					pack_len = len;
552 					boff = 0;
553 					bcopy_thresh = nxge_bcopy_thresh;
554 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
555 					    "==> nxge_start(7): > max NEW "
556 					    "bcopy thresh %d "
557 					    "pkt_len %d pack_len %d(next)",
558 					    bcopy_thresh,
559 					    pkt_len, pack_len));
560 				}
561 				last_bidx = i;
562 			}
563 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
564 			if ((boff == TX_PKT_HEADER_SIZE) && (nmblks == 1)) {
565 				hdrp = (p_tx_pkt_header_t)kaddr;
566 				header_set = B_TRUE;
567 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
568 				    "==> nxge_start(7_x2): "
569 				    "pkt_len %d pack_len %d (new hdrp $%p)",
570 				    pkt_len, pack_len, hdrp));
571 			}
572 			tx_msg_p->flags.dma_type = USE_BCOPY;
573 			kaddr += boff;
574 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(8): "
575 			    "USE BCOPY: before bcopy "
576 			    "DESC IOADDR $%p entry %d "
577 			    "bcopy packets %d "
578 			    "bcopy kaddr $%p "
579 			    "bcopy ioaddr (SAD) $%p "
580 			    "bcopy clen %d "
581 			    "bcopy boff %d",
582 			    DMA_COMMON_IOADDR(desc_area), i,
583 			    tdc_stats->tx_hdr_pkts,
584 			    kaddr,
585 			    dma_ioaddr,
586 			    clen,
587 			    boff));
588 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
589 			    "1USE BCOPY: "));
590 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
591 			    "2USE BCOPY: "));
592 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
593 			    "last USE BCOPY: copy from b_rptr $%p "
594 			    "to KADDR $%p (len %d offset %d",
595 			    b_rptr, kaddr, len, boff));
596 
597 			bcopy(b_rptr, kaddr, len);
598 
599 #ifdef	NXGE_DEBUG
600 			dump_len = (len > 128) ? 128: len;
601 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
602 			    "==> nxge_start: dump packets "
603 			    "(After BCOPY len %d)"
604 			    "(b_rptr $%p): %s", len, nmp->b_rptr,
605 			    nxge_dump_packet((char *)nmp->b_rptr,
606 			    dump_len)));
607 #endif
608 
609 			dma_handle = tx_msg_p->buf_dma_handle;
610 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
611 			(void) ddi_dma_sync(dma_handle,
612 			    i * nxge_bcopy_thresh, nxge_bcopy_thresh,
613 			    DDI_DMA_SYNC_FORDEV);
614 			clen = len + boff;
615 			tdc_stats->tx_hdr_pkts++;
616 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(9): "
617 			    "USE BCOPY: "
618 			    "DESC IOADDR $%p entry %d "
619 			    "bcopy packets %d "
620 			    "bcopy kaddr $%p "
621 			    "bcopy ioaddr (SAD) $%p "
622 			    "bcopy clen %d "
623 			    "bcopy boff %d",
624 			    DMA_COMMON_IOADDR(desc_area),
625 			    i,
626 			    tdc_stats->tx_hdr_pkts,
627 			    kaddr,
628 			    dma_ioaddr,
629 			    clen,
630 			    boff));
631 		} else {
632 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(12): "
633 			    "USE DVMA: len %d", len));
634 			tx_msg_p->flags.dma_type = USE_DMA;
635 			dma_flags = DDI_DMA_WRITE;
636 			if (len < nxge_dma_stream_thresh) {
637 				dma_flags |= DDI_DMA_CONSISTENT;
638 			} else {
639 				dma_flags |= DDI_DMA_STREAMING;
640 			}
641 
642 			dma_handle = tx_msg_p->dma_handle;
643 			dma_status = ddi_dma_addr_bind_handle(dma_handle, NULL,
644 			    (caddr_t)b_rptr, len, dma_flags,
645 			    DDI_DMA_DONTWAIT, NULL,
646 			    &dma_cookie, &ncookies);
647 			if (dma_status == DDI_DMA_MAPPED) {
648 				dma_ioaddr = dma_cookie.dmac_laddress;
649 				len = (int)dma_cookie.dmac_size;
650 				clen = (uint32_t)dma_cookie.dmac_size;
651 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
652 				    "==> nxge_start(12_1): "
653 				    "USE DVMA: len %d clen %d "
654 				    "ngathers %d",
655 				    len, clen,
656 				    ngathers));
657 #if defined(__i386)
658 				npi_desc_handle.regp = (uint32_t)tx_desc_p;
659 #else
660 				npi_desc_handle.regp = (uint64_t)tx_desc_p;
661 #endif
662 				while (ncookies > 1) {
663 					ngathers++;
664 					/*
665 					 * this is the fix for multiple
666 					 * cookies, which are basically
667 					 * a descriptor entry, we don't set
668 					 * SOP bit as well as related fields
669 					 */
670 
671 					(void) npi_txdma_desc_gather_set(
672 					    npi_desc_handle,
673 					    &tx_desc,
674 					    (ngathers -1),
675 					    mark_mode,
676 					    ngathers,
677 					    dma_ioaddr,
678 					    clen);
679 
680 					tx_msg_p->tx_msg_size = clen;
681 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
682 					    "==> nxge_start:  DMA "
683 					    "ncookie %d "
684 					    "ngathers %d "
685 					    "dma_ioaddr $%p len %d"
686 					    "desc $%p descp $%p (%d)",
687 					    ncookies,
688 					    ngathers,
689 					    dma_ioaddr, clen,
690 					    *tx_desc_p, tx_desc_p, i));
691 
692 					ddi_dma_nextcookie(dma_handle,
693 					    &dma_cookie);
694 					dma_ioaddr =
695 					    dma_cookie.dmac_laddress;
696 
697 					len = (int)dma_cookie.dmac_size;
698 					clen = (uint32_t)dma_cookie.dmac_size;
699 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
700 					    "==> nxge_start(12_2): "
701 					    "USE DVMA: len %d clen %d ",
702 					    len, clen));
703 
704 					i = TXDMA_DESC_NEXT_INDEX(i, 1,
705 					    tx_ring_p->tx_wrap_mask);
706 					tx_desc_p = &tx_desc_ring_vp[i];
707 
708 #if defined(__i386)
709 					npi_desc_handle.regp =
710 					    (uint32_t)tx_desc_p;
711 #else
712 					npi_desc_handle.regp =
713 					    (uint64_t)tx_desc_p;
714 #endif
715 					tx_msg_p = &tx_msg_ring[i];
716 					tx_msg_p->flags.dma_type = USE_NONE;
717 					tx_desc.value = 0;
718 
719 					ncookies--;
720 				}
721 				tdc_stats->tx_ddi_pkts++;
722 				NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start:"
723 				    "DMA: ddi packets %d",
724 				    tdc_stats->tx_ddi_pkts));
725 			} else {
726 				NXGE_ERROR_MSG((nxgep, NXGE_ERR_CTL,
727 				    "dma mapping failed for %d "
728 				    "bytes addr $%p flags %x (%d)",
729 				    len, b_rptr, status, status));
730 				good_packet = B_FALSE;
731 				tdc_stats->tx_dma_bind_fail++;
732 				tx_msg_p->flags.dma_type = USE_NONE;
733 				if (is_lso) {
734 					mp = nmp;
735 					goto nxge_start_fail_lso;
736 				} else {
737 					status = 1;
738 					goto nxge_start_fail2;
739 				}
740 			}
741 		} /* ddi dvma */
742 
743 		if (is_lso) {
744 			nmp_lso_save = nmp;
745 		}
746 		nmp = nmp->b_cont;
747 nxge_start_control_header_only:
748 #if defined(__i386)
749 		npi_desc_handle.regp = (uint32_t)tx_desc_p;
750 #else
751 		npi_desc_handle.regp = (uint64_t)tx_desc_p;
752 #endif
753 		ngathers++;
754 
755 		if (ngathers == 1) {
756 #ifdef	NXGE_DEBUG
757 			save_desc_p = &sop_tx_desc;
758 #endif
759 			sop_tx_desc_p = &sop_tx_desc;
760 			sop_tx_desc_p->value = 0;
761 			sop_tx_desc_p->bits.hdw.tr_len = clen;
762 			sop_tx_desc_p->bits.hdw.sad = dma_ioaddr >> 32;
763 			sop_tx_desc_p->bits.ldw.sad = dma_ioaddr & 0xffffffff;
764 		} else {
765 #ifdef	NXGE_DEBUG
766 			save_desc_p = &tx_desc;
767 #endif
768 			tmp_desc_p = &tx_desc;
769 			tmp_desc_p->value = 0;
770 			tmp_desc_p->bits.hdw.tr_len = clen;
771 			tmp_desc_p->bits.hdw.sad = dma_ioaddr >> 32;
772 			tmp_desc_p->bits.ldw.sad = dma_ioaddr & 0xffffffff;
773 
774 			tx_desc_p->value = tmp_desc_p->value;
775 		}
776 
777 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(13): "
778 		    "Desc_entry %d ngathers %d "
779 		    "desc_vp $%p tx_desc_p $%p "
780 		    "len %d clen %d pkt_len %d pack_len %d nmblks %d "
781 		    "dma_ioaddr (SAD) $%p mark %d",
782 		    i, ngathers,
783 		    tx_desc_ring_vp, tx_desc_p,
784 		    len, clen, pkt_len, pack_len, nmblks,
785 		    dma_ioaddr, mark_mode));
786 
787 #ifdef NXGE_DEBUG
788 		npi_desc_handle.nxgep = nxgep;
789 		npi_desc_handle.function.function = nxgep->function_num;
790 		npi_desc_handle.function.instance = nxgep->instance;
791 		sad = (save_desc_p->value & TX_PKT_DESC_SAD_MASK);
792 		xfer_len = ((save_desc_p->value & TX_PKT_DESC_TR_LEN_MASK) >>
793 		    TX_PKT_DESC_TR_LEN_SHIFT);
794 
795 
796 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "\n\t: value 0x%llx\n"
797 		    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\t"
798 		    "mark %d sop %d\n",
799 		    save_desc_p->value,
800 		    sad,
801 		    save_desc_p->bits.hdw.tr_len,
802 		    xfer_len,
803 		    save_desc_p->bits.hdw.num_ptr,
804 		    save_desc_p->bits.hdw.mark,
805 		    save_desc_p->bits.hdw.sop));
806 
807 		npi_txdma_dump_desc_one(npi_desc_handle, NULL, i);
808 #endif
809 
810 		tx_msg_p->tx_msg_size = clen;
811 		i = TXDMA_DESC_NEXT_INDEX(i, 1, tx_ring_p->tx_wrap_mask);
812 		if (ngathers > nxge_tx_max_gathers) {
813 			good_packet = B_FALSE;
814 			mac_hcksum_get(mp, &start_offset,
815 			    &stuff_offset, &end_offset, &value,
816 			    &cksum_flags);
817 
818 			NXGE_DEBUG_MSG((NULL, TX_CTL,
819 			    "==> nxge_start(14): pull msg - "
820 			    "len %d pkt_len %d ngathers %d",
821 			    len, pkt_len, ngathers));
822 
823 			/*
824 			 * Just give up on this packet.
825 			 */
826 			if (is_lso) {
827 				mp = nmp_lso_save;
828 				goto nxge_start_fail_lso;
829 			}
830 			status = 0;
831 			goto nxge_start_fail2;
832 		}
833 	} /* while (nmp) */
834 
835 	tx_msg_p->tx_message = mp;
836 	tx_desc_p = &tx_desc_ring_vp[sop_index];
837 #if defined(__i386)
838 	npi_desc_handle.regp = (uint32_t)tx_desc_p;
839 #else
840 	npi_desc_handle.regp = (uint64_t)tx_desc_p;
841 #endif
842 
843 	pkthdrp = (p_tx_pkt_hdr_all_t)hdrp;
844 	pkthdrp->reserved = 0;
845 	hdrp->value = 0;
846 	bcopy(&tmp_hdrp, hdrp, sizeof (tx_pkt_header_t));
847 
848 	if (pkt_len > NXGE_MTU_DEFAULT_MAX) {
849 		tdc_stats->tx_jumbo_pkts++;
850 	}
851 
852 	min_len = (ETHERMIN + TX_PKT_HEADER_SIZE + (npads * 2));
853 	if (pkt_len < min_len) {
854 		/* Assume we use bcopy to premapped buffers */
855 		kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
856 		NXGE_DEBUG_MSG((NULL, TX_CTL,
857 		    "==> nxge_start(14-1): < (msg_min + 16)"
858 		    "len %d pkt_len %d min_len %d bzero %d ngathers %d",
859 		    len, pkt_len, min_len, (min_len - pkt_len), ngathers));
860 		bzero((kaddr + pkt_len), (min_len - pkt_len));
861 		pkt_len = tx_msg_p->tx_msg_size = min_len;
862 
863 		sop_tx_desc_p->bits.hdw.tr_len = min_len;
864 
865 		NXGE_MEM_PIO_WRITE64(npi_desc_handle, sop_tx_desc_p->value);
866 		tx_desc_p->value = sop_tx_desc_p->value;
867 
868 		NXGE_DEBUG_MSG((NULL, TX_CTL,
869 		    "==> nxge_start(14-2): < msg_min - "
870 		    "len %d pkt_len %d min_len %d ngathers %d",
871 		    len, pkt_len, min_len, ngathers));
872 	}
873 
874 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: cksum_flags 0x%x ",
875 	    cksum_flags));
876 	{
877 		uint64_t	tmp_len;
878 
879 		/* pkt_len already includes 16 + paddings!! */
880 		/* Update the control header length */
881 		tot_xfer_len = (pkt_len - TX_PKT_HEADER_SIZE);
882 		tmp_len = hdrp->value |
883 		    (tot_xfer_len << TX_PKT_HEADER_TOT_XFER_LEN_SHIFT);
884 
885 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
886 		    "==> nxge_start(15_x1): setting SOP "
887 		    "tot_xfer_len 0x%llx (%d) pkt_len %d tmp_len "
888 		    "0x%llx hdrp->value 0x%llx",
889 		    tot_xfer_len, tot_xfer_len, pkt_len,
890 		    tmp_len, hdrp->value));
891 #if defined(_BIG_ENDIAN)
892 		hdrp->value = ddi_swap64(tmp_len);
893 #else
894 		hdrp->value = tmp_len;
895 #endif
896 		NXGE_DEBUG_MSG((nxgep,
897 		    TX_CTL, "==> nxge_start(15_x2): setting SOP "
898 		    "after SWAP: tot_xfer_len 0x%llx pkt_len %d "
899 		    "tmp_len 0x%llx hdrp->value 0x%llx",
900 		    tot_xfer_len, pkt_len,
901 		    tmp_len, hdrp->value));
902 	}
903 
904 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(15): setting SOP "
905 	    "wr_index %d "
906 	    "tot_xfer_len (%d) pkt_len %d npads %d",
907 	    sop_index,
908 	    tot_xfer_len, pkt_len,
909 	    npads));
910 
911 	sop_tx_desc_p->bits.hdw.sop = 1;
912 	sop_tx_desc_p->bits.hdw.mark = mark_mode;
913 	sop_tx_desc_p->bits.hdw.num_ptr = ngathers;
914 
915 	NXGE_MEM_PIO_WRITE64(npi_desc_handle, sop_tx_desc_p->value);
916 
917 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(16): set SOP done"));
918 
919 #ifdef NXGE_DEBUG
920 	npi_desc_handle.nxgep = nxgep;
921 	npi_desc_handle.function.function = nxgep->function_num;
922 	npi_desc_handle.function.instance = nxgep->instance;
923 
924 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "\n\t: value 0x%llx\n"
925 	    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\tmark %d sop %d\n",
926 	    save_desc_p->value,
927 	    sad,
928 	    save_desc_p->bits.hdw.tr_len,
929 	    xfer_len,
930 	    save_desc_p->bits.hdw.num_ptr,
931 	    save_desc_p->bits.hdw.mark,
932 	    save_desc_p->bits.hdw.sop));
933 	(void) npi_txdma_dump_desc_one(npi_desc_handle, NULL, sop_index);
934 
935 	dump_len = (pkt_len > 128) ? 128: pkt_len;
936 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
937 	    "==> nxge_start: dump packets(17) (after sop set, len "
938 	    " (len/dump_len/pkt_len/tot_xfer_len) %d/%d/%d/%d):\n"
939 	    "ptr $%p: %s", len, dump_len, pkt_len, tot_xfer_len,
940 	    (char *)hdrp,
941 	    nxge_dump_packet((char *)hdrp, dump_len)));
942 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
943 	    "==> nxge_start(18): TX desc sync: sop_index %d",
944 	    sop_index));
945 #endif
946 
947 	if ((ngathers == 1) || tx_ring_p->wr_index < i) {
948 		(void) ddi_dma_sync(tx_desc_dma_handle,
949 		    sop_index * sizeof (tx_desc_t),
950 		    ngathers * sizeof (tx_desc_t),
951 		    DDI_DMA_SYNC_FORDEV);
952 
953 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(19): sync 1 "
954 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
955 		    "pkt_len %d ngathers %d sop_index %d\n",
956 		    stuff_offset, start_offset,
957 		    pkt_len, ngathers, sop_index));
958 	} else { /* more than one descriptor and wrap around */
959 		uint32_t nsdescs = tx_ring_p->tx_ring_size - sop_index;
960 		(void) ddi_dma_sync(tx_desc_dma_handle,
961 		    sop_index * sizeof (tx_desc_t),
962 		    nsdescs * sizeof (tx_desc_t),
963 		    DDI_DMA_SYNC_FORDEV);
964 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(20): sync 1 "
965 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
966 		    "pkt_len %d ngathers %d sop_index %d\n",
967 		    stuff_offset, start_offset,
968 		    pkt_len, ngathers, sop_index));
969 
970 		(void) ddi_dma_sync(tx_desc_dma_handle,
971 		    0,
972 		    (ngathers - nsdescs) * sizeof (tx_desc_t),
973 		    DDI_DMA_SYNC_FORDEV);
974 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(21): sync 2 "
975 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
976 		    "pkt_len %d ngathers %d sop_index %d\n",
977 		    stuff_offset, start_offset,
978 		    pkt_len, ngathers, sop_index));
979 	}
980 
981 	tail_index = tx_ring_p->wr_index;
982 	tail_wrap = tx_ring_p->wr_index_wrap;
983 
984 	tx_ring_p->wr_index = i;
985 	if (tx_ring_p->wr_index <= tail_index) {
986 		tx_ring_p->wr_index_wrap = ((tail_wrap == B_TRUE) ?
987 		    B_FALSE : B_TRUE);
988 	}
989 
990 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: TX kick: "
991 	    "channel %d wr_index %d wrap %d ngathers %d desc_pend %d",
992 	    tx_ring_p->tdc,
993 	    tx_ring_p->wr_index,
994 	    tx_ring_p->wr_index_wrap,
995 	    ngathers,
996 	    tx_ring_p->descs_pending));
997 
998 	if (is_lso) {
999 		lso_ngathers += ngathers;
1000 		if (mp_chain != NULL) {
1001 			mp = mp_chain;
1002 			mp_chain = mp_chain->b_next;
1003 			mp->b_next = NULL;
1004 			if (nxge_lso_kick_cnt == lso_ngathers) {
1005 				tx_ring_p->descs_pending += lso_ngathers;
1006 				{
1007 					tx_ring_kick_t		kick;
1008 
1009 					kick.value = 0;
1010 					kick.bits.ldw.wrap =
1011 					    tx_ring_p->wr_index_wrap;
1012 					kick.bits.ldw.tail =
1013 					    (uint16_t)tx_ring_p->wr_index;
1014 
1015 					/* Kick the Transmit kick register */
1016 					TXDMA_REG_WRITE64(
1017 					    NXGE_DEV_NPI_HANDLE(nxgep),
1018 					    TX_RING_KICK_REG,
1019 					    (uint8_t)tx_ring_p->tdc,
1020 					    kick.value);
1021 					tdc_stats->tx_starts++;
1022 
1023 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
1024 					    "==> nxge_start: more LSO: "
1025 					    "LSO_CNT %d",
1026 					    lso_ngathers));
1027 				}
1028 				lso_ngathers = 0;
1029 				ngathers = 0;
1030 				cur_index_lso = sop_index = tx_ring_p->wr_index;
1031 				lso_tail_wrap = tx_ring_p->wr_index_wrap;
1032 			}
1033 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
1034 			    "==> nxge_start: lso again: "
1035 			    "lso_gathers %d ngathers %d cur_index_lso %d "
1036 			    "wr_index %d sop_index %d",
1037 			    lso_ngathers, ngathers, cur_index_lso,
1038 			    tx_ring_p->wr_index, sop_index));
1039 
1040 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
1041 			    "==> nxge_start: next : count %d",
1042 			    lso_ngathers));
1043 			lso_again = B_TRUE;
1044 			goto start_again;
1045 		}
1046 		ngathers = lso_ngathers;
1047 	}
1048 
1049 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: TX KICKING: "));
1050 
1051 	{
1052 		tx_ring_kick_t		kick;
1053 
1054 		kick.value = 0;
1055 		kick.bits.ldw.wrap = tx_ring_p->wr_index_wrap;
1056 		kick.bits.ldw.tail = (uint16_t)tx_ring_p->wr_index;
1057 
1058 		/* Kick start the Transmit kick register */
1059 		TXDMA_REG_WRITE64(NXGE_DEV_NPI_HANDLE(nxgep),
1060 		    TX_RING_KICK_REG,
1061 		    (uint8_t)tx_ring_p->tdc,
1062 		    kick.value);
1063 	}
1064 
1065 	tx_ring_p->descs_pending += ngathers;
1066 	tdc_stats->tx_starts++;
1067 
1068 	if (isLDOMservice(nxgep)) {
1069 		tx_ring_p->tx_ring_busy = B_FALSE;
1070 		if (tx_ring_p->tx_ring_offline) {
1071 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
1072 			    NXGE_TX_RING_OFFLINED);
1073 		}
1074 	}
1075 
1076 	MUTEX_EXIT(&tx_ring_p->lock);
1077 
1078 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "<== nxge_start"));
1079 	return (status);
1080 
1081 nxge_start_fail_lso:
1082 	status = 0;
1083 	good_packet = B_FALSE;
1084 	if (mp != NULL)
1085 		freemsg(mp);
1086 	if (mp_chain != NULL)
1087 		freemsgchain(mp_chain);
1088 
1089 	if (!lso_again && !ngathers) {
1090 		if (isLDOMservice(nxgep)) {
1091 			tx_ring_p->tx_ring_busy = B_FALSE;
1092 			if (tx_ring_p->tx_ring_offline) {
1093 				(void) atomic_swap_32(
1094 				    &tx_ring_p->tx_ring_offline,
1095 				    NXGE_TX_RING_OFFLINED);
1096 			}
1097 		}
1098 
1099 		MUTEX_EXIT(&tx_ring_p->lock);
1100 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
1101 		    "==> nxge_start: lso exit (nothing changed)"));
1102 		goto nxge_start_fail1;
1103 	}
1104 
1105 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
1106 	    "==> nxge_start (channel %d): before lso "
1107 	    "lso_gathers %d ngathers %d cur_index_lso %d "
1108 	    "wr_index %d sop_index %d lso_again %d",
1109 	    tx_ring_p->tdc,
1110 	    lso_ngathers, ngathers, cur_index_lso,
1111 	    tx_ring_p->wr_index, sop_index, lso_again));
1112 
1113 	if (lso_again) {
1114 		lso_ngathers += ngathers;
1115 		ngathers = lso_ngathers;
1116 		sop_index = cur_index_lso;
1117 		tx_ring_p->wr_index = sop_index;
1118 		tx_ring_p->wr_index_wrap = lso_tail_wrap;
1119 	}
1120 
1121 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
1122 	    "==> nxge_start (channel %d): after lso "
1123 	    "lso_gathers %d ngathers %d cur_index_lso %d "
1124 	    "wr_index %d sop_index %d lso_again %d",
1125 	    tx_ring_p->tdc,
1126 	    lso_ngathers, ngathers, cur_index_lso,
1127 	    tx_ring_p->wr_index, sop_index, lso_again));
1128 
1129 nxge_start_fail2:
1130 	if (good_packet == B_FALSE) {
1131 		cur_index = sop_index;
1132 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: clean up"));
1133 		for (i = 0; i < ngathers; i++) {
1134 			tx_desc_p = &tx_desc_ring_vp[cur_index];
1135 #if defined(__i386)
1136 			npi_handle.regp = (uint32_t)tx_desc_p;
1137 #else
1138 			npi_handle.regp = (uint64_t)tx_desc_p;
1139 #endif
1140 			tx_msg_p = &tx_msg_ring[cur_index];
1141 			(void) npi_txdma_desc_set_zero(npi_handle, 1);
1142 			if (tx_msg_p->flags.dma_type == USE_DVMA) {
1143 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
1144 				    "tx_desc_p = %X index = %d",
1145 				    tx_desc_p, tx_ring_p->rd_index));
1146 				(void) dvma_unload(tx_msg_p->dvma_handle,
1147 				    0, -1);
1148 				tx_msg_p->dvma_handle = NULL;
1149 				if (tx_ring_p->dvma_wr_index ==
1150 				    tx_ring_p->dvma_wrap_mask)
1151 					tx_ring_p->dvma_wr_index = 0;
1152 				else
1153 					tx_ring_p->dvma_wr_index++;
1154 				tx_ring_p->dvma_pending--;
1155 			} else if (tx_msg_p->flags.dma_type == USE_DMA) {
1156 				if (ddi_dma_unbind_handle(
1157 				    tx_msg_p->dma_handle)) {
1158 					cmn_err(CE_WARN, "!nxge_start: "
1159 					    "ddi_dma_unbind_handle failed");
1160 				}
1161 			}
1162 			tx_msg_p->flags.dma_type = USE_NONE;
1163 			cur_index = TXDMA_DESC_NEXT_INDEX(cur_index, 1,
1164 			    tx_ring_p->tx_wrap_mask);
1165 
1166 		}
1167 	}
1168 
1169 	if (isLDOMservice(nxgep)) {
1170 		tx_ring_p->tx_ring_busy = B_FALSE;
1171 		if (tx_ring_p->tx_ring_offline) {
1172 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
1173 			    NXGE_TX_RING_OFFLINED);
1174 		}
1175 	}
1176 
1177 	MUTEX_EXIT(&tx_ring_p->lock);
1178 
1179 nxge_start_fail1:
1180 	/* Add FMA to check the access handle nxge_hregh */
1181 
1182 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "<== nxge_start"));
1183 	return (status);
1184 }
1185 
1186 /* Software LSO starts here */
1187 static void
1188 nxge_hcksum_retrieve(mblk_t *mp,
1189     uint32_t *start, uint32_t *stuff, uint32_t *end,
1190     uint32_t *value, uint32_t *flags)
1191 {
1192 	if (mp->b_datap->db_type == M_DATA) {
1193 		if (flags != NULL) {
1194 			*flags = DB_CKSUMFLAGS(mp) & (HCK_IPV4_HDRCKSUM |
1195 			    HCK_PARTIALCKSUM | HCK_FULLCKSUM |
1196 			    HCK_FULLCKSUM_OK);
1197 			if ((*flags & (HCK_PARTIALCKSUM |
1198 			    HCK_FULLCKSUM)) != 0) {
1199 				if (value != NULL)
1200 					*value = (uint32_t)DB_CKSUM16(mp);
1201 				if ((*flags & HCK_PARTIALCKSUM) != 0) {
1202 					if (start != NULL)
1203 						*start =
1204 						    (uint32_t)DB_CKSUMSTART(mp);
1205 					if (stuff != NULL)
1206 						*stuff =
1207 						    (uint32_t)DB_CKSUMSTUFF(mp);
1208 					if (end != NULL)
1209 						*end =
1210 						    (uint32_t)DB_CKSUMEND(mp);
1211 				}
1212 			}
1213 		}
1214 	}
1215 }
1216 
1217 static void
1218 nxge_lso_info_get(mblk_t *mp, uint32_t *mss, uint32_t *flags)
1219 {
1220 	ASSERT(DB_TYPE(mp) == M_DATA);
1221 
1222 	*mss = 0;
1223 	if (flags != NULL) {
1224 		*flags = DB_CKSUMFLAGS(mp) & HW_LSO;
1225 		if ((*flags != 0) && (mss != NULL)) {
1226 			*mss = (uint32_t)DB_LSOMSS(mp);
1227 		}
1228 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1229 		    "==> nxge_lso_info_get(flag !=NULL): mss %d *flags 0x%x",
1230 		    *mss, *flags));
1231 	}
1232 
1233 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1234 	    "<== nxge_lso_info_get: mss %d", *mss));
1235 }
1236 
1237 /*
1238  * Do Soft LSO on the oversized packet.
1239  *
1240  * 1. Create a chain of message for headers.
1241  * 2. Fill up header messages with proper information.
1242  * 3. Copy Eithernet, IP, and TCP headers from the original message to
1243  *    each new message with necessary adjustments.
1244  *    * Unchange the ethernet header for DIX frames. (by default)
1245  *    * IP Total Length field is updated to MSS or less(only for the last one).
1246  *    * IP Identification value is incremented by one for each packet.
1247  *    * TCP sequence Number is recalculated according to the payload length.
1248  *    * Set FIN and/or PSH flags for the *last* packet if applied.
1249  *    * TCP partial Checksum
1250  * 4. Update LSO information in the first message header.
1251  * 5. Release the original message header.
1252  */
1253 static mblk_t *
1254 nxge_do_softlso(mblk_t *mp, uint32_t mss)
1255 {
1256 	uint32_t	hckflags;
1257 	int		pktlen;
1258 	int		hdrlen;
1259 	int		segnum;
1260 	int		i;
1261 	struct ether_vlan_header *evh;
1262 	int		ehlen, iphlen, tcphlen;
1263 	struct ip	*oiph, *niph;
1264 	struct tcphdr *otcph, *ntcph;
1265 	int		available, len, left;
1266 	uint16_t	ip_id;
1267 	uint32_t	tcp_seq;
1268 #ifdef __sparc
1269 	uint32_t	tcp_seq_tmp;
1270 #endif
1271 	mblk_t		*datamp;
1272 	uchar_t		*rptr;
1273 	mblk_t		*nmp;
1274 	mblk_t		*cmp;
1275 	mblk_t		*mp_chain;
1276 	boolean_t do_cleanup = B_FALSE;
1277 	t_uscalar_t start_offset = 0;
1278 	t_uscalar_t stuff_offset = 0;
1279 	t_uscalar_t value = 0;
1280 	uint16_t	l4_len;
1281 	ipaddr_t	src, dst;
1282 	uint32_t	cksum, sum, l4cksum;
1283 
1284 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1285 	    "==> nxge_do_softlso"));
1286 	/*
1287 	 * check the length of LSO packet payload and calculate the number of
1288 	 * segments to be generated.
1289 	 */
1290 	pktlen = msgsize(mp);
1291 	evh = (struct ether_vlan_header *)mp->b_rptr;
1292 
1293 	/* VLAN? */
1294 	if (evh->ether_tpid == htons(ETHERTYPE_VLAN))
1295 		ehlen = sizeof (struct ether_vlan_header);
1296 	else
1297 		ehlen = sizeof (struct ether_header);
1298 	oiph = (struct ip *)(mp->b_rptr + ehlen);
1299 	iphlen = oiph->ip_hl * 4;
1300 	otcph = (struct tcphdr *)(mp->b_rptr + ehlen + iphlen);
1301 	tcphlen = otcph->th_off * 4;
1302 
1303 	l4_len = pktlen - ehlen - iphlen;
1304 
1305 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1306 	    "==> nxge_do_softlso: mss %d oiph $%p "
1307 	    "original ip_sum oiph->ip_sum 0x%x "
1308 	    "original tcp_sum otcph->th_sum 0x%x "
1309 	    "oiph->ip_len %d pktlen %d ehlen %d "
1310 	    "l4_len %d (0x%x) ip_len - iphlen %d ",
1311 	    mss,
1312 	    oiph,
1313 	    oiph->ip_sum,
1314 	    otcph->th_sum,
1315 	    ntohs(oiph->ip_len), pktlen,
1316 	    ehlen,
1317 	    l4_len,
1318 	    l4_len,
1319 	    ntohs(oiph->ip_len) - iphlen));
1320 
1321 	/* IPv4 + TCP */
1322 	if (!(oiph->ip_v == IPV4_VERSION)) {
1323 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1324 		    "<== nxge_do_softlso: not IPV4 "
1325 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1326 		    ntohs(oiph->ip_len), pktlen, ehlen,
1327 		    tcphlen));
1328 		freemsg(mp);
1329 		return (NULL);
1330 	}
1331 
1332 	if (!(oiph->ip_p == IPPROTO_TCP)) {
1333 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1334 		    "<== nxge_do_softlso: not TCP "
1335 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1336 		    ntohs(oiph->ip_len), pktlen, ehlen,
1337 		    tcphlen));
1338 		freemsg(mp);
1339 		return (NULL);
1340 	}
1341 
1342 	if (!(ntohs(oiph->ip_len) == pktlen - ehlen)) {
1343 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1344 		    "<== nxge_do_softlso: len not matched  "
1345 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1346 		    ntohs(oiph->ip_len), pktlen, ehlen,
1347 		    tcphlen));
1348 		freemsg(mp);
1349 		return (NULL);
1350 	}
1351 
1352 	otcph = (struct tcphdr *)(mp->b_rptr + ehlen + iphlen);
1353 	tcphlen = otcph->th_off * 4;
1354 
1355 	/* TCP flags can not include URG, RST, or SYN */
1356 	VERIFY((otcph->th_flags & (TH_SYN | TH_RST | TH_URG)) == 0);
1357 
1358 	hdrlen = ehlen + iphlen + tcphlen;
1359 
1360 	VERIFY(MBLKL(mp) >= hdrlen);
1361 
1362 	if (MBLKL(mp) > hdrlen) {
1363 		datamp = mp;
1364 		rptr = mp->b_rptr + hdrlen;
1365 	} else { /* = */
1366 		datamp = mp->b_cont;
1367 		rptr = datamp->b_rptr;
1368 	}
1369 
1370 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1371 	    "nxge_do_softlso: otcph $%p pktlen: %d, "
1372 	    "hdrlen %d ehlen %d iphlen %d tcphlen %d "
1373 	    "mblkl(mp): %d, mblkl(datamp): %d",
1374 	    otcph,
1375 	    pktlen, hdrlen, ehlen, iphlen, tcphlen,
1376 	    (int)MBLKL(mp), (int)MBLKL(datamp)));
1377 
1378 	hckflags = 0;
1379 	nxge_hcksum_retrieve(mp,
1380 	    &start_offset, &stuff_offset, &value, NULL, &hckflags);
1381 
1382 	dst = oiph->ip_dst.s_addr;
1383 	src = oiph->ip_src.s_addr;
1384 
1385 	cksum = (dst >> 16) + (dst & 0xFFFF) +
1386 	    (src >> 16) + (src & 0xFFFF);
1387 	l4cksum = cksum + IP_TCP_CSUM_COMP;
1388 
1389 	sum = l4_len + l4cksum;
1390 	sum = (sum & 0xFFFF) + (sum >> 16);
1391 
1392 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1393 	    "==> nxge_do_softlso: dst 0x%x src 0x%x sum 0x%x ~new 0x%x "
1394 	    "hckflags 0x%x start_offset %d stuff_offset %d "
1395 	    "value (original) 0x%x th_sum 0x%x "
1396 	    "pktlen %d l4_len %d (0x%x) "
1397 	    "MBLKL(mp): %d, MBLKL(datamp): %d dump header %s",
1398 	    dst, src,
1399 	    (sum & 0xffff), (~sum & 0xffff),
1400 	    hckflags, start_offset, stuff_offset,
1401 	    value, otcph->th_sum,
1402 	    pktlen,
1403 	    l4_len,
1404 	    l4_len,
1405 	    ntohs(oiph->ip_len) - (int)MBLKL(mp),
1406 	    (int)MBLKL(datamp),
1407 	    nxge_dump_packet((char *)evh, 12)));
1408 
1409 	/*
1410 	 * Start to process.
1411 	 */
1412 	available = pktlen - hdrlen;
1413 	segnum = (available - 1) / mss + 1;
1414 
1415 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1416 	    "==> nxge_do_softlso: pktlen %d "
1417 	    "MBLKL(mp): %d, MBLKL(datamp): %d "
1418 	    "available %d mss %d segnum %d",
1419 	    pktlen, (int)MBLKL(mp), (int)MBLKL(datamp),
1420 	    available,
1421 	    mss,
1422 	    segnum));
1423 
1424 	VERIFY(segnum >= 2);
1425 
1426 	/*
1427 	 * Try to pre-allocate all header messages
1428 	 */
1429 	mp_chain = NULL;
1430 	for (i = 0; i < segnum; i++) {
1431 		if ((nmp = allocb(hdrlen, 0)) == NULL) {
1432 			/* Clean up the mp_chain */
1433 			while (mp_chain != NULL) {
1434 				nmp = mp_chain;
1435 				mp_chain = mp_chain->b_next;
1436 				freemsg(nmp);
1437 			}
1438 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1439 			    "<== nxge_do_softlso: "
1440 			    "Could not allocate enough messages for headers!"));
1441 			freemsg(mp);
1442 			return (NULL);
1443 		}
1444 		nmp->b_next = mp_chain;
1445 		mp_chain = nmp;
1446 
1447 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1448 		    "==> nxge_do_softlso: "
1449 		    "mp $%p nmp $%p mp_chain $%p mp_chain->b_next $%p",
1450 		    mp, nmp, mp_chain, mp_chain->b_next));
1451 	}
1452 
1453 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1454 	    "==> nxge_do_softlso: mp $%p nmp $%p mp_chain $%p",
1455 	    mp, nmp, mp_chain));
1456 
1457 	/*
1458 	 * Associate payload with new packets
1459 	 */
1460 	cmp = mp_chain;
1461 	left = available;
1462 	while (cmp != NULL) {
1463 		nmp = dupb(datamp);
1464 		if (nmp == NULL) {
1465 			do_cleanup = B_TRUE;
1466 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1467 			    "==>nxge_do_softlso: "
1468 			    "Can not dupb(datamp), have to do clean up"));
1469 			goto cleanup_allocated_msgs;
1470 		}
1471 
1472 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1473 		    "==> nxge_do_softlso: (loop) before mp $%p cmp $%p "
1474 		    "dupb nmp $%p len %d left %d msd %d ",
1475 		    mp, cmp, nmp, len, left, mss));
1476 
1477 		cmp->b_cont = nmp;
1478 		nmp->b_rptr = rptr;
1479 		len = (left < mss) ? left : mss;
1480 		left -= len;
1481 
1482 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1483 		    "==> nxge_do_softlso: (loop) after mp $%p cmp $%p "
1484 		    "dupb nmp $%p len %d left %d mss %d ",
1485 		    mp, cmp, nmp, len, left, mss));
1486 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1487 		    "nxge_do_softlso: before available: %d, "
1488 		    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1489 		    available, left, len, segnum, (int)MBLKL(nmp)));
1490 
1491 		len -= MBLKL(nmp);
1492 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1493 		    "nxge_do_softlso: after available: %d, "
1494 		    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1495 		    available, left, len, segnum, (int)MBLKL(nmp)));
1496 
1497 		while (len > 0) {
1498 			mblk_t *mmp = NULL;
1499 
1500 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1501 			    "nxge_do_softlso: (4) len > 0 available: %d, "
1502 			    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1503 			    available, left, len, segnum, (int)MBLKL(nmp)));
1504 
1505 			if (datamp->b_cont != NULL) {
1506 				datamp = datamp->b_cont;
1507 				rptr = datamp->b_rptr;
1508 				mmp = dupb(datamp);
1509 				if (mmp == NULL) {
1510 					do_cleanup = B_TRUE;
1511 					NXGE_DEBUG_MSG((NULL, TX_CTL,
1512 					    "==> nxge_do_softlso: "
1513 					    "Can not dupb(datamp) (1), :"
1514 					    "have to do clean up"));
1515 					NXGE_DEBUG_MSG((NULL, TX_CTL,
1516 					    "==> nxge_do_softlso: "
1517 					    "available: %d, left: %d, "
1518 					    "len: %d, MBLKL(nmp): %d",
1519 					    available, left, len,
1520 					    (int)MBLKL(nmp)));
1521 					goto cleanup_allocated_msgs;
1522 				}
1523 			} else {
1524 				NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1525 				    "==> nxge_do_softlso: "
1526 				    "(1)available: %d, left: %d, "
1527 				    "len: %d, MBLKL(nmp): %d",
1528 				    available, left, len,
1529 				    (int)MBLKL(nmp)));
1530 				cmn_err(CE_PANIC,
1531 				    "==> nxge_do_softlso: "
1532 				    "Pointers must have been corrupted!\n"
1533 				    "datamp: $%p, nmp: $%p, rptr: $%p",
1534 				    (void *)datamp,
1535 				    (void *)nmp,
1536 				    (void *)rptr);
1537 			}
1538 			nmp->b_cont = mmp;
1539 			nmp = mmp;
1540 			len -= MBLKL(nmp);
1541 		}
1542 		if (len < 0) {
1543 			nmp->b_wptr += len;
1544 			rptr = nmp->b_wptr;
1545 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1546 			    "(5) len < 0 (less than 0)"
1547 			    "available: %d, left: %d, len: %d, MBLKL(nmp): %d",
1548 			    available, left, len, (int)MBLKL(nmp)));
1549 
1550 		} else if (len == 0) {
1551 			if (datamp->b_cont != NULL) {
1552 				NXGE_DEBUG_MSG((NULL, TX_CTL,
1553 				    "(5) len == 0"
1554 				    "available: %d, left: %d, len: %d, "
1555 				    "MBLKL(nmp): %d",
1556 				    available, left, len, (int)MBLKL(nmp)));
1557 				datamp = datamp->b_cont;
1558 				rptr = datamp->b_rptr;
1559 			} else {
1560 				NXGE_DEBUG_MSG((NULL, TX_CTL,
1561 				    "(6)available b_cont == NULL : %d, "
1562 				    "left: %d, len: %d, MBLKL(nmp): %d",
1563 				    available, left, len, (int)MBLKL(nmp)));
1564 
1565 				VERIFY(cmp->b_next == NULL);
1566 				VERIFY(left == 0);
1567 				break; /* Done! */
1568 			}
1569 		}
1570 		cmp = cmp->b_next;
1571 
1572 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1573 		    "(7) do_softlso: "
1574 		    "next mp in mp_chain available len != 0 : %d, "
1575 		    "left: %d, len: %d, MBLKL(nmp): %d",
1576 		    available, left, len, (int)MBLKL(nmp)));
1577 	}
1578 
1579 	/*
1580 	 * From now, start to fill up all headers for the first message
1581 	 * Hardware checksum flags need to be updated separately for FULLCKSUM
1582 	 * and PARTIALCKSUM cases. For full checksum, copy the original flags
1583 	 * into every new packet is enough. But for HCK_PARTIALCKSUM, all
1584 	 * required fields need to be updated properly.
1585 	 */
1586 	nmp = mp_chain;
1587 	bcopy(mp->b_rptr, nmp->b_rptr, hdrlen);
1588 	nmp->b_wptr = nmp->b_rptr + hdrlen;
1589 	niph = (struct ip *)(nmp->b_rptr + ehlen);
1590 	niph->ip_len = htons(mss + iphlen + tcphlen);
1591 	ip_id = ntohs(niph->ip_id);
1592 	ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1593 #ifdef __sparc
1594 	bcopy((char *)&ntcph->th_seq, &tcp_seq_tmp, 4);
1595 	tcp_seq = ntohl(tcp_seq_tmp);
1596 #else
1597 	tcp_seq = ntohl(ntcph->th_seq);
1598 #endif
1599 
1600 	ntcph->th_flags &= ~(TH_FIN | TH_PUSH | TH_RST);
1601 
1602 	DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1603 	DB_CKSUMSTART(nmp) = start_offset;
1604 	DB_CKSUMSTUFF(nmp) = stuff_offset;
1605 
1606 	/* calculate IP checksum and TCP pseudo header checksum */
1607 	niph->ip_sum = 0;
1608 	niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1609 
1610 	l4_len = mss + tcphlen;
1611 	sum = htons(l4_len) + l4cksum;
1612 	sum = (sum & 0xFFFF) + (sum >> 16);
1613 	ntcph->th_sum = (sum & 0xffff);
1614 
1615 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1616 	    "==> nxge_do_softlso: first mp $%p (mp_chain $%p) "
1617 	    "mss %d pktlen %d l4_len %d (0x%x) "
1618 	    "MBLKL(mp): %d, MBLKL(datamp): %d "
1619 	    "ip_sum 0x%x "
1620 	    "th_sum 0x%x sum 0x%x ) "
1621 	    "dump first ip->tcp %s",
1622 	    nmp, mp_chain,
1623 	    mss,
1624 	    pktlen,
1625 	    l4_len,
1626 	    l4_len,
1627 	    (int)MBLKL(mp), (int)MBLKL(datamp),
1628 	    niph->ip_sum,
1629 	    ntcph->th_sum,
1630 	    sum,
1631 	    nxge_dump_packet((char *)niph, 52)));
1632 
1633 	cmp = nmp;
1634 	while ((nmp = nmp->b_next)->b_next != NULL) {
1635 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1636 		    "==>nxge_do_softlso: middle l4_len %d ", l4_len));
1637 		bcopy(cmp->b_rptr, nmp->b_rptr, hdrlen);
1638 		nmp->b_wptr = nmp->b_rptr + hdrlen;
1639 		niph = (struct ip *)(nmp->b_rptr + ehlen);
1640 		niph->ip_id = htons(++ip_id);
1641 		niph->ip_len = htons(mss + iphlen + tcphlen);
1642 		ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1643 		tcp_seq += mss;
1644 
1645 		ntcph->th_flags &= ~(TH_FIN | TH_PUSH | TH_RST | TH_URG);
1646 
1647 #ifdef __sparc
1648 		tcp_seq_tmp = htonl(tcp_seq);
1649 		bcopy(&tcp_seq_tmp, (char *)&ntcph->th_seq, 4);
1650 #else
1651 		ntcph->th_seq = htonl(tcp_seq);
1652 #endif
1653 		DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1654 		DB_CKSUMSTART(nmp) = start_offset;
1655 		DB_CKSUMSTUFF(nmp) = stuff_offset;
1656 
1657 		/* calculate IP checksum and TCP pseudo header checksum */
1658 		niph->ip_sum = 0;
1659 		niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1660 		ntcph->th_sum = (sum & 0xffff);
1661 
1662 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1663 		    "==> nxge_do_softlso: middle ip_sum 0x%x "
1664 		    "th_sum 0x%x "
1665 		    " mp $%p (mp_chain $%p) pktlen %d "
1666 		    "MBLKL(mp): %d, MBLKL(datamp): %d ",
1667 		    niph->ip_sum,
1668 		    ntcph->th_sum,
1669 		    nmp, mp_chain,
1670 		    pktlen, (int)MBLKL(mp), (int)MBLKL(datamp)));
1671 	}
1672 
1673 	/* Last segment */
1674 	/*
1675 	 * Set FIN and/or PSH flags if present only in the last packet.
1676 	 * The ip_len could be different from prior packets.
1677 	 */
1678 	bcopy(cmp->b_rptr, nmp->b_rptr, hdrlen);
1679 	nmp->b_wptr = nmp->b_rptr + hdrlen;
1680 	niph = (struct ip *)(nmp->b_rptr + ehlen);
1681 	niph->ip_id = htons(++ip_id);
1682 	niph->ip_len = htons(msgsize(nmp->b_cont) + iphlen + tcphlen);
1683 	ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1684 	tcp_seq += mss;
1685 #ifdef __sparc
1686 	tcp_seq_tmp = htonl(tcp_seq);
1687 	bcopy(&tcp_seq_tmp, (char *)&ntcph->th_seq, 4);
1688 #else
1689 	ntcph->th_seq = htonl(tcp_seq);
1690 #endif
1691 	ntcph->th_flags = (otcph->th_flags & ~TH_URG);
1692 
1693 	DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1694 	DB_CKSUMSTART(nmp) = start_offset;
1695 	DB_CKSUMSTUFF(nmp) = stuff_offset;
1696 
1697 	/* calculate IP checksum and TCP pseudo header checksum */
1698 	niph->ip_sum = 0;
1699 	niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1700 
1701 	l4_len = ntohs(niph->ip_len) - iphlen;
1702 	sum = htons(l4_len) + l4cksum;
1703 	sum = (sum & 0xFFFF) + (sum >> 16);
1704 	ntcph->th_sum = (sum & 0xffff);
1705 
1706 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1707 	    "==> nxge_do_softlso: last next "
1708 	    "niph->ip_sum 0x%x "
1709 	    "ntcph->th_sum 0x%x sum 0x%x "
1710 	    "dump last ip->tcp %s "
1711 	    "cmp $%p mp $%p (mp_chain $%p) pktlen %d (0x%x) "
1712 	    "l4_len %d (0x%x) "
1713 	    "MBLKL(mp): %d, MBLKL(datamp): %d ",
1714 	    niph->ip_sum,
1715 	    ntcph->th_sum, sum,
1716 	    nxge_dump_packet((char *)niph, 52),
1717 	    cmp, nmp, mp_chain,
1718 	    pktlen, pktlen,
1719 	    l4_len,
1720 	    l4_len,
1721 	    (int)MBLKL(mp), (int)MBLKL(datamp)));
1722 
1723 cleanup_allocated_msgs:
1724 	if (do_cleanup) {
1725 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1726 		    "==> nxge_do_softlso: "
1727 		    "Failed allocating messages, "
1728 		    "have to clean up and fail!"));
1729 		while (mp_chain != NULL) {
1730 			nmp = mp_chain;
1731 			mp_chain = mp_chain->b_next;
1732 			freemsg(nmp);
1733 		}
1734 	}
1735 	/*
1736 	 * We're done here, so just free the original message and return the
1737 	 * new message chain, that could be NULL if failed, back to the caller.
1738 	 */
1739 	freemsg(mp);
1740 
1741 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1742 	    "<== nxge_do_softlso:mp_chain $%p", mp_chain));
1743 	return (mp_chain);
1744 }
1745 
1746 /*
1747  * Will be called before NIC driver do further operation on the message.
1748  * The input message may include LSO information, if so, go to softlso logic
1749  * to eliminate the oversized LSO packet for the incapable underlying h/w.
1750  * The return could be the same non-LSO message or a message chain for LSO case.
1751  *
1752  * The driver needs to call this function per packet and process the whole chain
1753  * if applied.
1754  */
1755 static mblk_t *
1756 nxge_lso_eliminate(mblk_t *mp)
1757 {
1758 	uint32_t lsoflags;
1759 	uint32_t mss;
1760 
1761 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1762 	    "==>nxge_lso_eliminate:"));
1763 	nxge_lso_info_get(mp, &mss, &lsoflags);
1764 
1765 	if (lsoflags & HW_LSO) {
1766 		mblk_t *nmp;
1767 
1768 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1769 		    "==>nxge_lso_eliminate:"
1770 		    "HW_LSO:mss %d mp $%p",
1771 		    mss, mp));
1772 		if ((nmp = nxge_do_softlso(mp, mss)) != NULL) {
1773 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1774 			    "<== nxge_lso_eliminate: "
1775 			    "LSO: nmp not NULL nmp $%p mss %d mp $%p",
1776 			    nmp, mss, mp));
1777 			return (nmp);
1778 		} else {
1779 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1780 			    "<== nxge_lso_eliminate_ "
1781 			    "LSO: failed nmp NULL nmp $%p mss %d mp $%p",
1782 			    nmp, mss, mp));
1783 			return (NULL);
1784 		}
1785 	}
1786 
1787 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1788 	    "<== nxge_lso_eliminate"));
1789 	return (mp);
1790 }
1791 
1792 static uint32_t
1793 nxge_csgen(uint16_t *adr, int len)
1794 {
1795 	int		i, odd;
1796 	uint32_t	sum = 0;
1797 	uint32_t	c = 0;
1798 
1799 	odd = len % 2;
1800 	for (i = 0; i < (len / 2); i++) {
1801 		sum += (adr[i] & 0xffff);
1802 	}
1803 	if (odd) {
1804 		sum += adr[len / 2] & 0xff00;
1805 	}
1806 	while ((c = ((sum & 0xffff0000) >> 16)) != 0) {
1807 		sum &= 0xffff;
1808 		sum += c;
1809 	}
1810 	return (~sum & 0xffff);
1811 }
1812