xref: /illumos-gate/usr/src/uts/common/io/nxge/nxge_send.c (revision 79a6379db826cfcc73d1752a7cff267414431708)
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 = NULL;
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 = NULL;
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 = NULL;
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) atomic_cas_32(
377 				    (uint32_t *)&tx_ring_p->queueing, 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 		npi_desc_handle.regp = (uint64_t)tx_desc_p;
487 		if (!header_set &&
488 		    ((!nxge_tx_use_bcopy && (len > TX_BCOPY_SIZE)) ||
489 		    (len >= bcopy_thresh))) {
490 			header_set = B_TRUE;
491 			bcopy_thresh += TX_PKT_HEADER_SIZE;
492 			boff = 0;
493 			pack_len = 0;
494 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
495 			hdrp = (p_tx_pkt_header_t)kaddr;
496 			clen = pkt_len;
497 			dma_handle = tx_msg_p->buf_dma_handle;
498 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
499 			(void) ddi_dma_sync(dma_handle,
500 			    i * nxge_bcopy_thresh, nxge_bcopy_thresh,
501 			    DDI_DMA_SYNC_FORDEV);
502 
503 			tx_msg_p->flags.dma_type = USE_BCOPY;
504 			goto nxge_start_control_header_only;
505 		}
506 
507 		pkt_len += len;
508 		pack_len += len;
509 
510 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(3): "
511 		    "desc entry %d "
512 		    "DESC IOADDR $%p "
513 		    "desc_vp $%p tx_desc_p $%p "
514 		    "desc_pp $%p tx_desc_pp $%p "
515 		    "len %d pkt_len %d pack_len %d",
516 		    i,
517 		    DMA_COMMON_IOADDR(desc_area),
518 		    tx_desc_ring_vp, tx_desc_p,
519 		    tx_desc_ring_pp, tx_desc_pp,
520 		    len, pkt_len, pack_len));
521 
522 		if (len < bcopy_thresh) {
523 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(4): "
524 			    "USE BCOPY: "));
525 			if (nxge_tx_tiny_pack) {
526 				uint32_t blst =
527 				    TXDMA_DESC_NEXT_INDEX(i, -1,
528 				    tx_ring_p->tx_wrap_mask);
529 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
530 				    "==> nxge_start(5): pack"));
531 				if ((pack_len <= bcopy_thresh) &&
532 				    (last_bidx == blst)) {
533 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
534 					    "==> nxge_start: pack(6) "
535 					    "(pkt_len %d pack_len %d)",
536 					    pkt_len, pack_len));
537 					i = blst;
538 					tx_desc_p = &tx_desc_ring_vp[i];
539 #ifdef	NXGE_DEBUG
540 					tx_desc_pp = &tx_desc_ring_pp[i];
541 #endif
542 					tx_msg_p = &tx_msg_ring[i];
543 					boff = pack_len - len;
544 					ngathers--;
545 				} else if (pack_len > bcopy_thresh &&
546 				    header_set) {
547 					pack_len = len;
548 					boff = 0;
549 					bcopy_thresh = nxge_bcopy_thresh;
550 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
551 					    "==> nxge_start(7): > max NEW "
552 					    "bcopy thresh %d "
553 					    "pkt_len %d pack_len %d(next)",
554 					    bcopy_thresh,
555 					    pkt_len, pack_len));
556 				}
557 				last_bidx = i;
558 			}
559 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
560 			if ((boff == TX_PKT_HEADER_SIZE) && (nmblks == 1)) {
561 				hdrp = (p_tx_pkt_header_t)kaddr;
562 				header_set = B_TRUE;
563 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
564 				    "==> nxge_start(7_x2): "
565 				    "pkt_len %d pack_len %d (new hdrp $%p)",
566 				    pkt_len, pack_len, hdrp));
567 			}
568 			tx_msg_p->flags.dma_type = USE_BCOPY;
569 			kaddr += boff;
570 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(8): "
571 			    "USE BCOPY: before bcopy "
572 			    "DESC IOADDR $%p entry %d "
573 			    "bcopy packets %d "
574 			    "bcopy kaddr $%p "
575 			    "bcopy ioaddr (SAD) $%p "
576 			    "bcopy clen %d "
577 			    "bcopy boff %d",
578 			    DMA_COMMON_IOADDR(desc_area), i,
579 			    tdc_stats->tx_hdr_pkts,
580 			    kaddr,
581 			    dma_ioaddr,
582 			    clen,
583 			    boff));
584 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
585 			    "1USE BCOPY: "));
586 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
587 			    "2USE BCOPY: "));
588 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
589 			    "last USE BCOPY: copy from b_rptr $%p "
590 			    "to KADDR $%p (len %d offset %d",
591 			    b_rptr, kaddr, len, boff));
592 
593 			bcopy(b_rptr, kaddr, len);
594 
595 #ifdef	NXGE_DEBUG
596 			dump_len = (len > 128) ? 128: len;
597 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
598 			    "==> nxge_start: dump packets "
599 			    "(After BCOPY len %d)"
600 			    "(b_rptr $%p): %s", len, nmp->b_rptr,
601 			    nxge_dump_packet((char *)nmp->b_rptr,
602 			    dump_len)));
603 #endif
604 
605 			dma_handle = tx_msg_p->buf_dma_handle;
606 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
607 			(void) ddi_dma_sync(dma_handle,
608 			    i * nxge_bcopy_thresh, nxge_bcopy_thresh,
609 			    DDI_DMA_SYNC_FORDEV);
610 			clen = len + boff;
611 			tdc_stats->tx_hdr_pkts++;
612 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(9): "
613 			    "USE BCOPY: "
614 			    "DESC IOADDR $%p entry %d "
615 			    "bcopy packets %d "
616 			    "bcopy kaddr $%p "
617 			    "bcopy ioaddr (SAD) $%p "
618 			    "bcopy clen %d "
619 			    "bcopy boff %d",
620 			    DMA_COMMON_IOADDR(desc_area),
621 			    i,
622 			    tdc_stats->tx_hdr_pkts,
623 			    kaddr,
624 			    dma_ioaddr,
625 			    clen,
626 			    boff));
627 		} else {
628 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(12): "
629 			    "USE DVMA: len %d", len));
630 			tx_msg_p->flags.dma_type = USE_DMA;
631 			dma_flags = DDI_DMA_WRITE;
632 			if (len < nxge_dma_stream_thresh) {
633 				dma_flags |= DDI_DMA_CONSISTENT;
634 			} else {
635 				dma_flags |= DDI_DMA_STREAMING;
636 			}
637 
638 			dma_handle = tx_msg_p->dma_handle;
639 			dma_status = ddi_dma_addr_bind_handle(dma_handle, NULL,
640 			    (caddr_t)b_rptr, len, dma_flags,
641 			    DDI_DMA_DONTWAIT, NULL,
642 			    &dma_cookie, &ncookies);
643 			if (dma_status == DDI_DMA_MAPPED) {
644 				dma_ioaddr = dma_cookie.dmac_laddress;
645 				len = (int)dma_cookie.dmac_size;
646 				clen = (uint32_t)dma_cookie.dmac_size;
647 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
648 				    "==> nxge_start(12_1): "
649 				    "USE DVMA: len %d clen %d "
650 				    "ngathers %d",
651 				    len, clen,
652 				    ngathers));
653 				npi_desc_handle.regp = (uint64_t)tx_desc_p;
654 				while (ncookies > 1) {
655 					ngathers++;
656 					/*
657 					 * this is the fix for multiple
658 					 * cookies, which are basically
659 					 * a descriptor entry, we don't set
660 					 * SOP bit as well as related fields
661 					 */
662 
663 					(void) npi_txdma_desc_gather_set(
664 					    npi_desc_handle,
665 					    &tx_desc,
666 					    (ngathers -1),
667 					    mark_mode,
668 					    ngathers,
669 					    dma_ioaddr,
670 					    clen);
671 
672 					tx_msg_p->tx_msg_size = clen;
673 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
674 					    "==> nxge_start:  DMA "
675 					    "ncookie %d "
676 					    "ngathers %d "
677 					    "dma_ioaddr $%p len %d"
678 					    "desc $%p descp $%p (%d)",
679 					    ncookies,
680 					    ngathers,
681 					    dma_ioaddr, clen,
682 					    *tx_desc_p, tx_desc_p, i));
683 
684 					ddi_dma_nextcookie(dma_handle,
685 					    &dma_cookie);
686 					dma_ioaddr =
687 					    dma_cookie.dmac_laddress;
688 
689 					len = (int)dma_cookie.dmac_size;
690 					clen = (uint32_t)dma_cookie.dmac_size;
691 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
692 					    "==> nxge_start(12_2): "
693 					    "USE DVMA: len %d clen %d ",
694 					    len, clen));
695 
696 					i = TXDMA_DESC_NEXT_INDEX(i, 1,
697 					    tx_ring_p->tx_wrap_mask);
698 					tx_desc_p = &tx_desc_ring_vp[i];
699 
700 					npi_desc_handle.regp =
701 					    (uint64_t)tx_desc_p;
702 					tx_msg_p = &tx_msg_ring[i];
703 					tx_msg_p->flags.dma_type = USE_NONE;
704 					tx_desc.value = 0;
705 
706 					ncookies--;
707 				}
708 				tdc_stats->tx_ddi_pkts++;
709 				NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start:"
710 				    "DMA: ddi packets %d",
711 				    tdc_stats->tx_ddi_pkts));
712 			} else {
713 				NXGE_ERROR_MSG((nxgep, NXGE_ERR_CTL,
714 				    "dma mapping failed for %d "
715 				    "bytes addr $%p flags %x (%d)",
716 				    len, b_rptr, status, status));
717 				good_packet = B_FALSE;
718 				tdc_stats->tx_dma_bind_fail++;
719 				tx_msg_p->flags.dma_type = USE_NONE;
720 				if (is_lso) {
721 					mp = nmp;
722 					goto nxge_start_fail_lso;
723 				} else {
724 					status = 1;
725 					goto nxge_start_fail2;
726 				}
727 			}
728 		} /* ddi dvma */
729 
730 		if (is_lso) {
731 			nmp_lso_save = nmp;
732 		}
733 		nmp = nmp->b_cont;
734 nxge_start_control_header_only:
735 		npi_desc_handle.regp = (uint64_t)tx_desc_p;
736 		ngathers++;
737 
738 		if (ngathers == 1) {
739 #ifdef	NXGE_DEBUG
740 			save_desc_p = &sop_tx_desc;
741 #endif
742 			sop_tx_desc_p = &sop_tx_desc;
743 			sop_tx_desc_p->value = 0;
744 			sop_tx_desc_p->bits.hdw.tr_len = clen;
745 			sop_tx_desc_p->bits.hdw.sad = dma_ioaddr >> 32;
746 			sop_tx_desc_p->bits.ldw.sad = dma_ioaddr & 0xffffffff;
747 		} else {
748 #ifdef	NXGE_DEBUG
749 			save_desc_p = &tx_desc;
750 #endif
751 			tmp_desc_p = &tx_desc;
752 			tmp_desc_p->value = 0;
753 			tmp_desc_p->bits.hdw.tr_len = clen;
754 			tmp_desc_p->bits.hdw.sad = dma_ioaddr >> 32;
755 			tmp_desc_p->bits.ldw.sad = dma_ioaddr & 0xffffffff;
756 
757 			tx_desc_p->value = tmp_desc_p->value;
758 		}
759 
760 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(13): "
761 		    "Desc_entry %d ngathers %d "
762 		    "desc_vp $%p tx_desc_p $%p "
763 		    "len %d clen %d pkt_len %d pack_len %d nmblks %d "
764 		    "dma_ioaddr (SAD) $%p mark %d",
765 		    i, ngathers,
766 		    tx_desc_ring_vp, tx_desc_p,
767 		    len, clen, pkt_len, pack_len, nmblks,
768 		    dma_ioaddr, mark_mode));
769 
770 #ifdef NXGE_DEBUG
771 		npi_desc_handle.nxgep = nxgep;
772 		npi_desc_handle.function.function = nxgep->function_num;
773 		npi_desc_handle.function.instance = nxgep->instance;
774 		sad = (save_desc_p->value & TX_PKT_DESC_SAD_MASK);
775 		xfer_len = ((save_desc_p->value & TX_PKT_DESC_TR_LEN_MASK) >>
776 		    TX_PKT_DESC_TR_LEN_SHIFT);
777 
778 
779 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "\n\t: value 0x%llx\n"
780 		    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\t"
781 		    "mark %d sop %d\n",
782 		    save_desc_p->value,
783 		    sad,
784 		    save_desc_p->bits.hdw.tr_len,
785 		    xfer_len,
786 		    save_desc_p->bits.hdw.num_ptr,
787 		    save_desc_p->bits.hdw.mark,
788 		    save_desc_p->bits.hdw.sop));
789 
790 		npi_txdma_dump_desc_one(npi_desc_handle, NULL, i);
791 #endif
792 
793 		tx_msg_p->tx_msg_size = clen;
794 		i = TXDMA_DESC_NEXT_INDEX(i, 1, tx_ring_p->tx_wrap_mask);
795 		if (ngathers > nxge_tx_max_gathers) {
796 			good_packet = B_FALSE;
797 			mac_hcksum_get(mp, &start_offset,
798 			    &stuff_offset, &end_offset, &value,
799 			    &cksum_flags);
800 
801 			NXGE_DEBUG_MSG((NULL, TX_CTL,
802 			    "==> nxge_start(14): pull msg - "
803 			    "len %d pkt_len %d ngathers %d",
804 			    len, pkt_len, ngathers));
805 
806 			/*
807 			 * Just give up on this packet.
808 			 */
809 			if (is_lso) {
810 				mp = nmp_lso_save;
811 				goto nxge_start_fail_lso;
812 			}
813 			status = 0;
814 			goto nxge_start_fail2;
815 		}
816 	} /* while (nmp) */
817 
818 	tx_msg_p->tx_message = mp;
819 	tx_desc_p = &tx_desc_ring_vp[sop_index];
820 	npi_desc_handle.regp = (uint64_t)tx_desc_p;
821 
822 	pkthdrp = (p_tx_pkt_hdr_all_t)hdrp;
823 	pkthdrp->reserved = 0;
824 	hdrp->value = 0;
825 	bcopy(&tmp_hdrp, hdrp, sizeof (tx_pkt_header_t));
826 
827 	if (pkt_len > NXGE_MTU_DEFAULT_MAX) {
828 		tdc_stats->tx_jumbo_pkts++;
829 	}
830 
831 	min_len = (ETHERMIN + TX_PKT_HEADER_SIZE + (npads * 2));
832 	if (pkt_len < min_len) {
833 		/* Assume we use bcopy to premapped buffers */
834 		kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
835 		NXGE_DEBUG_MSG((NULL, TX_CTL,
836 		    "==> nxge_start(14-1): < (msg_min + 16)"
837 		    "len %d pkt_len %d min_len %d bzero %d ngathers %d",
838 		    len, pkt_len, min_len, (min_len - pkt_len), ngathers));
839 		bzero((kaddr + pkt_len), (min_len - pkt_len));
840 		pkt_len = tx_msg_p->tx_msg_size = min_len;
841 
842 		sop_tx_desc_p->bits.hdw.tr_len = min_len;
843 
844 		NXGE_MEM_PIO_WRITE64(npi_desc_handle, sop_tx_desc_p->value);
845 		tx_desc_p->value = sop_tx_desc_p->value;
846 
847 		NXGE_DEBUG_MSG((NULL, TX_CTL,
848 		    "==> nxge_start(14-2): < msg_min - "
849 		    "len %d pkt_len %d min_len %d ngathers %d",
850 		    len, pkt_len, min_len, ngathers));
851 	}
852 
853 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: cksum_flags 0x%x ",
854 	    cksum_flags));
855 	{
856 		uint64_t	tmp_len;
857 
858 		/* pkt_len already includes 16 + paddings!! */
859 		/* Update the control header length */
860 		tot_xfer_len = (pkt_len - TX_PKT_HEADER_SIZE);
861 		tmp_len = hdrp->value |
862 		    (tot_xfer_len << TX_PKT_HEADER_TOT_XFER_LEN_SHIFT);
863 
864 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
865 		    "==> nxge_start(15_x1): setting SOP "
866 		    "tot_xfer_len 0x%llx (%d) pkt_len %d tmp_len "
867 		    "0x%llx hdrp->value 0x%llx",
868 		    tot_xfer_len, tot_xfer_len, pkt_len,
869 		    tmp_len, hdrp->value));
870 #if defined(_BIG_ENDIAN)
871 		hdrp->value = ddi_swap64(tmp_len);
872 #else
873 		hdrp->value = tmp_len;
874 #endif
875 		NXGE_DEBUG_MSG((nxgep,
876 		    TX_CTL, "==> nxge_start(15_x2): setting SOP "
877 		    "after SWAP: tot_xfer_len 0x%llx pkt_len %d "
878 		    "tmp_len 0x%llx hdrp->value 0x%llx",
879 		    tot_xfer_len, pkt_len,
880 		    tmp_len, hdrp->value));
881 	}
882 
883 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(15): setting SOP "
884 	    "wr_index %d "
885 	    "tot_xfer_len (%d) pkt_len %d npads %d",
886 	    sop_index,
887 	    tot_xfer_len, pkt_len,
888 	    npads));
889 
890 	sop_tx_desc_p->bits.hdw.sop = 1;
891 	sop_tx_desc_p->bits.hdw.mark = mark_mode;
892 	sop_tx_desc_p->bits.hdw.num_ptr = ngathers;
893 
894 	NXGE_MEM_PIO_WRITE64(npi_desc_handle, sop_tx_desc_p->value);
895 
896 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(16): set SOP done"));
897 
898 #ifdef NXGE_DEBUG
899 	npi_desc_handle.nxgep = nxgep;
900 	npi_desc_handle.function.function = nxgep->function_num;
901 	npi_desc_handle.function.instance = nxgep->instance;
902 
903 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "\n\t: value 0x%llx\n"
904 	    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\tmark %d sop %d\n",
905 	    save_desc_p->value,
906 	    sad,
907 	    save_desc_p->bits.hdw.tr_len,
908 	    xfer_len,
909 	    save_desc_p->bits.hdw.num_ptr,
910 	    save_desc_p->bits.hdw.mark,
911 	    save_desc_p->bits.hdw.sop));
912 	(void) npi_txdma_dump_desc_one(npi_desc_handle, NULL, sop_index);
913 
914 	dump_len = (pkt_len > 128) ? 128: pkt_len;
915 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
916 	    "==> nxge_start: dump packets(17) (after sop set, len "
917 	    " (len/dump_len/pkt_len/tot_xfer_len) %d/%d/%d/%d):\n"
918 	    "ptr $%p: %s", len, dump_len, pkt_len, tot_xfer_len,
919 	    (char *)hdrp,
920 	    nxge_dump_packet((char *)hdrp, dump_len)));
921 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
922 	    "==> nxge_start(18): TX desc sync: sop_index %d",
923 	    sop_index));
924 #endif
925 
926 	if ((ngathers == 1) || tx_ring_p->wr_index < i) {
927 		(void) ddi_dma_sync(tx_desc_dma_handle,
928 		    sop_index * sizeof (tx_desc_t),
929 		    ngathers * sizeof (tx_desc_t),
930 		    DDI_DMA_SYNC_FORDEV);
931 
932 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(19): sync 1 "
933 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
934 		    "pkt_len %d ngathers %d sop_index %d\n",
935 		    stuff_offset, start_offset,
936 		    pkt_len, ngathers, sop_index));
937 	} else { /* more than one descriptor and wrap around */
938 		uint32_t nsdescs = tx_ring_p->tx_ring_size - sop_index;
939 		(void) ddi_dma_sync(tx_desc_dma_handle,
940 		    sop_index * sizeof (tx_desc_t),
941 		    nsdescs * sizeof (tx_desc_t),
942 		    DDI_DMA_SYNC_FORDEV);
943 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(20): sync 1 "
944 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
945 		    "pkt_len %d ngathers %d sop_index %d\n",
946 		    stuff_offset, start_offset,
947 		    pkt_len, ngathers, sop_index));
948 
949 		(void) ddi_dma_sync(tx_desc_dma_handle,
950 		    0,
951 		    (ngathers - nsdescs) * sizeof (tx_desc_t),
952 		    DDI_DMA_SYNC_FORDEV);
953 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(21): sync 2 "
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 	}
959 
960 	tail_index = tx_ring_p->wr_index;
961 	tail_wrap = tx_ring_p->wr_index_wrap;
962 
963 	tx_ring_p->wr_index = i;
964 	if (tx_ring_p->wr_index <= tail_index) {
965 		tx_ring_p->wr_index_wrap = ((tail_wrap == B_TRUE) ?
966 		    B_FALSE : B_TRUE);
967 	}
968 
969 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: TX kick: "
970 	    "channel %d wr_index %d wrap %d ngathers %d desc_pend %d",
971 	    tx_ring_p->tdc,
972 	    tx_ring_p->wr_index,
973 	    tx_ring_p->wr_index_wrap,
974 	    ngathers,
975 	    tx_ring_p->descs_pending));
976 
977 	if (is_lso) {
978 		lso_ngathers += ngathers;
979 		if (mp_chain != NULL) {
980 			mp = mp_chain;
981 			mp_chain = mp_chain->b_next;
982 			mp->b_next = NULL;
983 			if (nxge_lso_kick_cnt == lso_ngathers) {
984 				tx_ring_p->descs_pending += lso_ngathers;
985 				{
986 					tx_ring_kick_t		kick;
987 
988 					kick.value = 0;
989 					kick.bits.ldw.wrap =
990 					    tx_ring_p->wr_index_wrap;
991 					kick.bits.ldw.tail =
992 					    (uint16_t)tx_ring_p->wr_index;
993 
994 					/* Kick the Transmit kick register */
995 					TXDMA_REG_WRITE64(
996 					    NXGE_DEV_NPI_HANDLE(nxgep),
997 					    TX_RING_KICK_REG,
998 					    (uint8_t)tx_ring_p->tdc,
999 					    kick.value);
1000 					tdc_stats->tx_starts++;
1001 
1002 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
1003 					    "==> nxge_start: more LSO: "
1004 					    "LSO_CNT %d",
1005 					    lso_ngathers));
1006 				}
1007 				lso_ngathers = 0;
1008 				ngathers = 0;
1009 				cur_index_lso = sop_index = tx_ring_p->wr_index;
1010 				lso_tail_wrap = tx_ring_p->wr_index_wrap;
1011 			}
1012 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
1013 			    "==> nxge_start: lso again: "
1014 			    "lso_gathers %d ngathers %d cur_index_lso %d "
1015 			    "wr_index %d sop_index %d",
1016 			    lso_ngathers, ngathers, cur_index_lso,
1017 			    tx_ring_p->wr_index, sop_index));
1018 
1019 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
1020 			    "==> nxge_start: next : count %d",
1021 			    lso_ngathers));
1022 			lso_again = B_TRUE;
1023 			goto start_again;
1024 		}
1025 		ngathers = lso_ngathers;
1026 	}
1027 
1028 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: TX KICKING: "));
1029 
1030 	{
1031 		tx_ring_kick_t		kick;
1032 
1033 		kick.value = 0;
1034 		kick.bits.ldw.wrap = tx_ring_p->wr_index_wrap;
1035 		kick.bits.ldw.tail = (uint16_t)tx_ring_p->wr_index;
1036 
1037 		/* Kick start the Transmit kick register */
1038 		TXDMA_REG_WRITE64(NXGE_DEV_NPI_HANDLE(nxgep),
1039 		    TX_RING_KICK_REG,
1040 		    (uint8_t)tx_ring_p->tdc,
1041 		    kick.value);
1042 	}
1043 
1044 	tx_ring_p->descs_pending += ngathers;
1045 	tdc_stats->tx_starts++;
1046 
1047 	if (isLDOMservice(nxgep)) {
1048 		tx_ring_p->tx_ring_busy = B_FALSE;
1049 		if (tx_ring_p->tx_ring_offline) {
1050 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
1051 			    NXGE_TX_RING_OFFLINED);
1052 		}
1053 	}
1054 
1055 	MUTEX_EXIT(&tx_ring_p->lock);
1056 
1057 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "<== nxge_start"));
1058 	return (status);
1059 
1060 nxge_start_fail_lso:
1061 	status = 0;
1062 	good_packet = B_FALSE;
1063 	if (mp != NULL)
1064 		freemsg(mp);
1065 	if (mp_chain != NULL)
1066 		freemsgchain(mp_chain);
1067 
1068 	if (!lso_again && !ngathers) {
1069 		if (isLDOMservice(nxgep)) {
1070 			tx_ring_p->tx_ring_busy = B_FALSE;
1071 			if (tx_ring_p->tx_ring_offline) {
1072 				(void) atomic_swap_32(
1073 				    &tx_ring_p->tx_ring_offline,
1074 				    NXGE_TX_RING_OFFLINED);
1075 			}
1076 		}
1077 
1078 		MUTEX_EXIT(&tx_ring_p->lock);
1079 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
1080 		    "==> nxge_start: lso exit (nothing changed)"));
1081 		goto nxge_start_fail1;
1082 	}
1083 
1084 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
1085 	    "==> nxge_start (channel %d): before lso "
1086 	    "lso_gathers %d ngathers %d cur_index_lso %d "
1087 	    "wr_index %d sop_index %d lso_again %d",
1088 	    tx_ring_p->tdc,
1089 	    lso_ngathers, ngathers, cur_index_lso,
1090 	    tx_ring_p->wr_index, sop_index, lso_again));
1091 
1092 	if (lso_again) {
1093 		lso_ngathers += ngathers;
1094 		ngathers = lso_ngathers;
1095 		sop_index = cur_index_lso;
1096 		tx_ring_p->wr_index = sop_index;
1097 		tx_ring_p->wr_index_wrap = lso_tail_wrap;
1098 	}
1099 
1100 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
1101 	    "==> nxge_start (channel %d): after lso "
1102 	    "lso_gathers %d ngathers %d cur_index_lso %d "
1103 	    "wr_index %d sop_index %d lso_again %d",
1104 	    tx_ring_p->tdc,
1105 	    lso_ngathers, ngathers, cur_index_lso,
1106 	    tx_ring_p->wr_index, sop_index, lso_again));
1107 
1108 nxge_start_fail2:
1109 	if (good_packet == B_FALSE) {
1110 		cur_index = sop_index;
1111 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: clean up"));
1112 		for (i = 0; i < ngathers; i++) {
1113 			tx_desc_p = &tx_desc_ring_vp[cur_index];
1114 			npi_handle.regp = (uint64_t)tx_desc_p;
1115 			tx_msg_p = &tx_msg_ring[cur_index];
1116 			(void) npi_txdma_desc_set_zero(npi_handle, 1);
1117 			if (tx_msg_p->flags.dma_type == USE_DVMA) {
1118 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
1119 				    "tx_desc_p = %X index = %d",
1120 				    tx_desc_p, tx_ring_p->rd_index));
1121 				(void) dvma_unload(tx_msg_p->dvma_handle,
1122 				    0, -1);
1123 				tx_msg_p->dvma_handle = NULL;
1124 				if (tx_ring_p->dvma_wr_index ==
1125 				    tx_ring_p->dvma_wrap_mask)
1126 					tx_ring_p->dvma_wr_index = 0;
1127 				else
1128 					tx_ring_p->dvma_wr_index++;
1129 				tx_ring_p->dvma_pending--;
1130 			} else if (tx_msg_p->flags.dma_type == USE_DMA) {
1131 				if (ddi_dma_unbind_handle(
1132 				    tx_msg_p->dma_handle)) {
1133 					cmn_err(CE_WARN, "!nxge_start: "
1134 					    "ddi_dma_unbind_handle failed");
1135 				}
1136 			}
1137 			tx_msg_p->flags.dma_type = USE_NONE;
1138 			cur_index = TXDMA_DESC_NEXT_INDEX(cur_index, 1,
1139 			    tx_ring_p->tx_wrap_mask);
1140 
1141 		}
1142 	}
1143 
1144 	if (isLDOMservice(nxgep)) {
1145 		tx_ring_p->tx_ring_busy = B_FALSE;
1146 		if (tx_ring_p->tx_ring_offline) {
1147 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
1148 			    NXGE_TX_RING_OFFLINED);
1149 		}
1150 	}
1151 
1152 	MUTEX_EXIT(&tx_ring_p->lock);
1153 
1154 nxge_start_fail1:
1155 	/* Add FMA to check the access handle nxge_hregh */
1156 
1157 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "<== nxge_start"));
1158 	return (status);
1159 }
1160 
1161 /* Software LSO starts here */
1162 static void
1163 nxge_hcksum_retrieve(mblk_t *mp,
1164     uint32_t *start, uint32_t *stuff, uint32_t *end,
1165     uint32_t *value, uint32_t *flags)
1166 {
1167 	if (mp->b_datap->db_type == M_DATA) {
1168 		if (flags != NULL) {
1169 			*flags = DB_CKSUMFLAGS(mp) & (HCK_IPV4_HDRCKSUM |
1170 			    HCK_PARTIALCKSUM | HCK_FULLCKSUM |
1171 			    HCK_FULLCKSUM_OK);
1172 			if ((*flags & (HCK_PARTIALCKSUM |
1173 			    HCK_FULLCKSUM)) != 0) {
1174 				if (value != NULL)
1175 					*value = (uint32_t)DB_CKSUM16(mp);
1176 				if ((*flags & HCK_PARTIALCKSUM) != 0) {
1177 					if (start != NULL)
1178 						*start =
1179 						    (uint32_t)DB_CKSUMSTART(mp);
1180 					if (stuff != NULL)
1181 						*stuff =
1182 						    (uint32_t)DB_CKSUMSTUFF(mp);
1183 					if (end != NULL)
1184 						*end =
1185 						    (uint32_t)DB_CKSUMEND(mp);
1186 				}
1187 			}
1188 		}
1189 	}
1190 }
1191 
1192 static void
1193 nxge_lso_info_get(mblk_t *mp, uint32_t *mss, uint32_t *flags)
1194 {
1195 	ASSERT(DB_TYPE(mp) == M_DATA);
1196 
1197 	if (mss == NULL || flags == NULL)
1198 		return;
1199 
1200 	*mss = 0;
1201 	if (flags != NULL) {
1202 		*flags = DB_CKSUMFLAGS(mp) & HW_LSO;
1203 		if (*flags != 0) {
1204 			*mss = (uint32_t)DB_LSOMSS(mp);
1205 		}
1206 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1207 		    "==> nxge_lso_info_get(flag !=NULL): mss %d *flags 0x%x",
1208 		    *mss, *flags));
1209 	}
1210 
1211 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1212 	    "<== nxge_lso_info_get: mss %d", *mss));
1213 }
1214 
1215 /*
1216  * Do Soft LSO on the oversized packet.
1217  *
1218  * 1. Create a chain of message for headers.
1219  * 2. Fill up header messages with proper information.
1220  * 3. Copy Eithernet, IP, and TCP headers from the original message to
1221  *    each new message with necessary adjustments.
1222  *    * Unchange the ethernet header for DIX frames. (by default)
1223  *    * IP Total Length field is updated to MSS or less(only for the last one).
1224  *    * IP Identification value is incremented by one for each packet.
1225  *    * TCP sequence Number is recalculated according to the payload length.
1226  *    * Set FIN and/or PSH flags for the *last* packet if applied.
1227  *    * TCP partial Checksum
1228  * 4. Update LSO information in the first message header.
1229  * 5. Release the original message header.
1230  */
1231 static mblk_t *
1232 nxge_do_softlso(mblk_t *mp, uint32_t mss)
1233 {
1234 	uint32_t	hckflags;
1235 	int		pktlen;
1236 	int		hdrlen;
1237 	int		segnum;
1238 	int		i;
1239 	struct ether_vlan_header *evh;
1240 	int		ehlen, iphlen, tcphlen;
1241 	struct ip	*oiph, *niph;
1242 	struct tcphdr *otcph, *ntcph;
1243 	int		available, len, left;
1244 	uint16_t	ip_id;
1245 	uint32_t	tcp_seq;
1246 #ifdef __sparc
1247 	uint32_t	tcp_seq_tmp;
1248 #endif
1249 	mblk_t		*datamp;
1250 	uchar_t		*rptr;
1251 	mblk_t		*nmp;
1252 	mblk_t		*cmp;
1253 	mblk_t		*mp_chain;
1254 	boolean_t do_cleanup = B_FALSE;
1255 	t_uscalar_t start_offset = 0;
1256 	t_uscalar_t stuff_offset = 0;
1257 	t_uscalar_t value = 0;
1258 	uint16_t	l4_len;
1259 	ipaddr_t	src, dst;
1260 	uint32_t	cksum, sum, l4cksum;
1261 
1262 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1263 	    "==> nxge_do_softlso"));
1264 	/*
1265 	 * check the length of LSO packet payload and calculate the number of
1266 	 * segments to be generated.
1267 	 */
1268 	pktlen = msgsize(mp);
1269 	evh = (struct ether_vlan_header *)mp->b_rptr;
1270 
1271 	/* VLAN? */
1272 	if (evh->ether_tpid == htons(ETHERTYPE_VLAN))
1273 		ehlen = sizeof (struct ether_vlan_header);
1274 	else
1275 		ehlen = sizeof (struct ether_header);
1276 	oiph = (struct ip *)(mp->b_rptr + ehlen);
1277 	iphlen = oiph->ip_hl * 4;
1278 	otcph = (struct tcphdr *)(mp->b_rptr + ehlen + iphlen);
1279 	tcphlen = otcph->th_off * 4;
1280 
1281 	l4_len = pktlen - ehlen - iphlen;
1282 
1283 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1284 	    "==> nxge_do_softlso: mss %d oiph $%p "
1285 	    "original ip_sum oiph->ip_sum 0x%x "
1286 	    "original tcp_sum otcph->th_sum 0x%x "
1287 	    "oiph->ip_len %d pktlen %d ehlen %d "
1288 	    "l4_len %d (0x%x) ip_len - iphlen %d ",
1289 	    mss,
1290 	    oiph,
1291 	    oiph->ip_sum,
1292 	    otcph->th_sum,
1293 	    ntohs(oiph->ip_len), pktlen,
1294 	    ehlen,
1295 	    l4_len,
1296 	    l4_len,
1297 	    ntohs(oiph->ip_len) - iphlen));
1298 
1299 	/* IPv4 + TCP */
1300 	if (!(oiph->ip_v == IPV4_VERSION)) {
1301 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1302 		    "<== nxge_do_softlso: not IPV4 "
1303 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1304 		    ntohs(oiph->ip_len), pktlen, ehlen,
1305 		    tcphlen));
1306 		freemsg(mp);
1307 		return (NULL);
1308 	}
1309 
1310 	if (!(oiph->ip_p == IPPROTO_TCP)) {
1311 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1312 		    "<== nxge_do_softlso: not TCP "
1313 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1314 		    ntohs(oiph->ip_len), pktlen, ehlen,
1315 		    tcphlen));
1316 		freemsg(mp);
1317 		return (NULL);
1318 	}
1319 
1320 	if (!(ntohs(oiph->ip_len) == pktlen - ehlen)) {
1321 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1322 		    "<== nxge_do_softlso: len not matched  "
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 	otcph = (struct tcphdr *)(mp->b_rptr + ehlen + iphlen);
1331 	tcphlen = otcph->th_off * 4;
1332 
1333 	/* TCP flags can not include URG, RST, or SYN */
1334 	VERIFY((otcph->th_flags & (TH_SYN | TH_RST | TH_URG)) == 0);
1335 
1336 	hdrlen = ehlen + iphlen + tcphlen;
1337 
1338 	VERIFY(MBLKL(mp) >= hdrlen);
1339 
1340 	if (MBLKL(mp) > hdrlen) {
1341 		datamp = mp;
1342 		rptr = mp->b_rptr + hdrlen;
1343 	} else { /* = */
1344 		datamp = mp->b_cont;
1345 		rptr = datamp->b_rptr;
1346 	}
1347 
1348 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1349 	    "nxge_do_softlso: otcph $%p pktlen: %d, "
1350 	    "hdrlen %d ehlen %d iphlen %d tcphlen %d "
1351 	    "mblkl(mp): %d, mblkl(datamp): %d",
1352 	    otcph,
1353 	    pktlen, hdrlen, ehlen, iphlen, tcphlen,
1354 	    (int)MBLKL(mp), (int)MBLKL(datamp)));
1355 
1356 	hckflags = 0;
1357 	nxge_hcksum_retrieve(mp,
1358 	    &start_offset, &stuff_offset, &value, NULL, &hckflags);
1359 
1360 	dst = oiph->ip_dst.s_addr;
1361 	src = oiph->ip_src.s_addr;
1362 
1363 	cksum = (dst >> 16) + (dst & 0xFFFF) +
1364 	    (src >> 16) + (src & 0xFFFF);
1365 	l4cksum = cksum + IP_TCP_CSUM_COMP;
1366 
1367 	sum = l4_len + l4cksum;
1368 	sum = (sum & 0xFFFF) + (sum >> 16);
1369 
1370 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1371 	    "==> nxge_do_softlso: dst 0x%x src 0x%x sum 0x%x ~new 0x%x "
1372 	    "hckflags 0x%x start_offset %d stuff_offset %d "
1373 	    "value (original) 0x%x th_sum 0x%x "
1374 	    "pktlen %d l4_len %d (0x%x) "
1375 	    "MBLKL(mp): %d, MBLKL(datamp): %d dump header %s",
1376 	    dst, src,
1377 	    (sum & 0xffff), (~sum & 0xffff),
1378 	    hckflags, start_offset, stuff_offset,
1379 	    value, otcph->th_sum,
1380 	    pktlen,
1381 	    l4_len,
1382 	    l4_len,
1383 	    ntohs(oiph->ip_len) - (int)MBLKL(mp),
1384 	    (int)MBLKL(datamp),
1385 	    nxge_dump_packet((char *)evh, 12)));
1386 
1387 	/*
1388 	 * Start to process.
1389 	 */
1390 	available = pktlen - hdrlen;
1391 	segnum = (available - 1) / mss + 1;
1392 
1393 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1394 	    "==> nxge_do_softlso: pktlen %d "
1395 	    "MBLKL(mp): %d, MBLKL(datamp): %d "
1396 	    "available %d mss %d segnum %d",
1397 	    pktlen, (int)MBLKL(mp), (int)MBLKL(datamp),
1398 	    available,
1399 	    mss,
1400 	    segnum));
1401 
1402 	VERIFY(segnum >= 2);
1403 
1404 	/*
1405 	 * Try to pre-allocate all header messages
1406 	 */
1407 	mp_chain = NULL;
1408 	for (i = 0; i < segnum; i++) {
1409 		if ((nmp = allocb(hdrlen, 0)) == NULL) {
1410 			/* Clean up the mp_chain */
1411 			while (mp_chain != NULL) {
1412 				nmp = mp_chain;
1413 				mp_chain = mp_chain->b_next;
1414 				freemsg(nmp);
1415 			}
1416 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1417 			    "<== nxge_do_softlso: "
1418 			    "Could not allocate enough messages for headers!"));
1419 			freemsg(mp);
1420 			return (NULL);
1421 		}
1422 		nmp->b_next = mp_chain;
1423 		mp_chain = nmp;
1424 
1425 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1426 		    "==> nxge_do_softlso: "
1427 		    "mp $%p nmp $%p mp_chain $%p mp_chain->b_next $%p",
1428 		    mp, nmp, mp_chain, mp_chain->b_next));
1429 	}
1430 
1431 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1432 	    "==> nxge_do_softlso: mp $%p nmp $%p mp_chain $%p",
1433 	    mp, nmp, mp_chain));
1434 
1435 	/*
1436 	 * Associate payload with new packets
1437 	 */
1438 	cmp = mp_chain;
1439 	left = available;
1440 	while (cmp != NULL) {
1441 		nmp = dupb(datamp);
1442 		if (nmp == NULL) {
1443 			do_cleanup = B_TRUE;
1444 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1445 			    "==>nxge_do_softlso: "
1446 			    "Can not dupb(datamp), have to do clean up"));
1447 			goto cleanup_allocated_msgs;
1448 		}
1449 
1450 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1451 		    "==> nxge_do_softlso: (loop) before mp $%p cmp $%p "
1452 		    "dupb nmp $%p len %d left %d msd %d ",
1453 		    mp, cmp, nmp, len, left, mss));
1454 
1455 		cmp->b_cont = nmp;
1456 		nmp->b_rptr = rptr;
1457 		len = (left < mss) ? left : mss;
1458 		left -= len;
1459 
1460 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1461 		    "==> nxge_do_softlso: (loop) after mp $%p cmp $%p "
1462 		    "dupb nmp $%p len %d left %d mss %d ",
1463 		    mp, cmp, nmp, len, left, mss));
1464 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1465 		    "nxge_do_softlso: before available: %d, "
1466 		    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1467 		    available, left, len, segnum, (int)MBLKL(nmp)));
1468 
1469 		len -= MBLKL(nmp);
1470 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1471 		    "nxge_do_softlso: after available: %d, "
1472 		    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1473 		    available, left, len, segnum, (int)MBLKL(nmp)));
1474 
1475 		while (len > 0) {
1476 			mblk_t *mmp = NULL;
1477 
1478 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1479 			    "nxge_do_softlso: (4) len > 0 available: %d, "
1480 			    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1481 			    available, left, len, segnum, (int)MBLKL(nmp)));
1482 
1483 			if (datamp->b_cont != NULL) {
1484 				datamp = datamp->b_cont;
1485 				rptr = datamp->b_rptr;
1486 				mmp = dupb(datamp);
1487 				if (mmp == NULL) {
1488 					do_cleanup = B_TRUE;
1489 					NXGE_DEBUG_MSG((NULL, TX_CTL,
1490 					    "==> nxge_do_softlso: "
1491 					    "Can not dupb(datamp) (1), :"
1492 					    "have to do clean up"));
1493 					NXGE_DEBUG_MSG((NULL, TX_CTL,
1494 					    "==> nxge_do_softlso: "
1495 					    "available: %d, left: %d, "
1496 					    "len: %d, MBLKL(nmp): %d",
1497 					    available, left, len,
1498 					    (int)MBLKL(nmp)));
1499 					goto cleanup_allocated_msgs;
1500 				}
1501 			} else {
1502 				NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1503 				    "==> nxge_do_softlso: "
1504 				    "(1)available: %d, left: %d, "
1505 				    "len: %d, MBLKL(nmp): %d",
1506 				    available, left, len,
1507 				    (int)MBLKL(nmp)));
1508 				cmn_err(CE_PANIC,
1509 				    "==> nxge_do_softlso: "
1510 				    "Pointers must have been corrupted!\n"
1511 				    "datamp: $%p, nmp: $%p, rptr: $%p",
1512 				    (void *)datamp,
1513 				    (void *)nmp,
1514 				    (void *)rptr);
1515 			}
1516 			nmp->b_cont = mmp;
1517 			nmp = mmp;
1518 			len -= MBLKL(nmp);
1519 		}
1520 		if (len < 0) {
1521 			nmp->b_wptr += len;
1522 			rptr = nmp->b_wptr;
1523 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1524 			    "(5) len < 0 (less than 0)"
1525 			    "available: %d, left: %d, len: %d, MBLKL(nmp): %d",
1526 			    available, left, len, (int)MBLKL(nmp)));
1527 
1528 		} else if (len == 0) {
1529 			if (datamp->b_cont != NULL) {
1530 				NXGE_DEBUG_MSG((NULL, TX_CTL,
1531 				    "(5) len == 0"
1532 				    "available: %d, left: %d, len: %d, "
1533 				    "MBLKL(nmp): %d",
1534 				    available, left, len, (int)MBLKL(nmp)));
1535 				datamp = datamp->b_cont;
1536 				rptr = datamp->b_rptr;
1537 			} else {
1538 				NXGE_DEBUG_MSG((NULL, TX_CTL,
1539 				    "(6)available b_cont == NULL : %d, "
1540 				    "left: %d, len: %d, MBLKL(nmp): %d",
1541 				    available, left, len, (int)MBLKL(nmp)));
1542 
1543 				VERIFY(cmp->b_next == NULL);
1544 				VERIFY(left == 0);
1545 				break; /* Done! */
1546 			}
1547 		}
1548 		cmp = cmp->b_next;
1549 
1550 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1551 		    "(7) do_softlso: "
1552 		    "next mp in mp_chain available len != 0 : %d, "
1553 		    "left: %d, len: %d, MBLKL(nmp): %d",
1554 		    available, left, len, (int)MBLKL(nmp)));
1555 	}
1556 
1557 	/*
1558 	 * From now, start to fill up all headers for the first message
1559 	 * Hardware checksum flags need to be updated separately for FULLCKSUM
1560 	 * and PARTIALCKSUM cases. For full checksum, copy the original flags
1561 	 * into every new packet is enough. But for HCK_PARTIALCKSUM, all
1562 	 * required fields need to be updated properly.
1563 	 */
1564 	nmp = mp_chain;
1565 	bcopy(mp->b_rptr, nmp->b_rptr, hdrlen);
1566 	nmp->b_wptr = nmp->b_rptr + hdrlen;
1567 	niph = (struct ip *)(nmp->b_rptr + ehlen);
1568 	niph->ip_len = htons(mss + iphlen + tcphlen);
1569 	ip_id = ntohs(niph->ip_id);
1570 	ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1571 #ifdef __sparc
1572 	bcopy((char *)&ntcph->th_seq, &tcp_seq_tmp, 4);
1573 	tcp_seq = ntohl(tcp_seq_tmp);
1574 #else
1575 	tcp_seq = ntohl(ntcph->th_seq);
1576 #endif
1577 
1578 	ntcph->th_flags &= ~(TH_FIN | TH_PUSH | TH_RST);
1579 
1580 	DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1581 	DB_CKSUMSTART(nmp) = start_offset;
1582 	DB_CKSUMSTUFF(nmp) = stuff_offset;
1583 
1584 	/* calculate IP checksum and TCP pseudo header checksum */
1585 	niph->ip_sum = 0;
1586 	niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1587 
1588 	l4_len = mss + tcphlen;
1589 	sum = htons(l4_len) + l4cksum;
1590 	sum = (sum & 0xFFFF) + (sum >> 16);
1591 	ntcph->th_sum = (sum & 0xffff);
1592 
1593 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1594 	    "==> nxge_do_softlso: first mp $%p (mp_chain $%p) "
1595 	    "mss %d pktlen %d l4_len %d (0x%x) "
1596 	    "MBLKL(mp): %d, MBLKL(datamp): %d "
1597 	    "ip_sum 0x%x "
1598 	    "th_sum 0x%x sum 0x%x ) "
1599 	    "dump first ip->tcp %s",
1600 	    nmp, mp_chain,
1601 	    mss,
1602 	    pktlen,
1603 	    l4_len,
1604 	    l4_len,
1605 	    (int)MBLKL(mp), (int)MBLKL(datamp),
1606 	    niph->ip_sum,
1607 	    ntcph->th_sum,
1608 	    sum,
1609 	    nxge_dump_packet((char *)niph, 52)));
1610 
1611 	cmp = nmp;
1612 	while ((nmp = nmp->b_next)->b_next != NULL) {
1613 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1614 		    "==>nxge_do_softlso: middle l4_len %d ", l4_len));
1615 		bcopy(cmp->b_rptr, nmp->b_rptr, hdrlen);
1616 		nmp->b_wptr = nmp->b_rptr + hdrlen;
1617 		niph = (struct ip *)(nmp->b_rptr + ehlen);
1618 		niph->ip_id = htons(++ip_id);
1619 		niph->ip_len = htons(mss + iphlen + tcphlen);
1620 		ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1621 		tcp_seq += mss;
1622 
1623 		ntcph->th_flags &= ~(TH_FIN | TH_PUSH | TH_RST | TH_URG);
1624 
1625 #ifdef __sparc
1626 		tcp_seq_tmp = htonl(tcp_seq);
1627 		bcopy(&tcp_seq_tmp, (char *)&ntcph->th_seq, 4);
1628 #else
1629 		ntcph->th_seq = htonl(tcp_seq);
1630 #endif
1631 		DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1632 		DB_CKSUMSTART(nmp) = start_offset;
1633 		DB_CKSUMSTUFF(nmp) = stuff_offset;
1634 
1635 		/* calculate IP checksum and TCP pseudo header checksum */
1636 		niph->ip_sum = 0;
1637 		niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1638 		ntcph->th_sum = (sum & 0xffff);
1639 
1640 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1641 		    "==> nxge_do_softlso: middle ip_sum 0x%x "
1642 		    "th_sum 0x%x "
1643 		    " mp $%p (mp_chain $%p) pktlen %d "
1644 		    "MBLKL(mp): %d, MBLKL(datamp): %d ",
1645 		    niph->ip_sum,
1646 		    ntcph->th_sum,
1647 		    nmp, mp_chain,
1648 		    pktlen, (int)MBLKL(mp), (int)MBLKL(datamp)));
1649 	}
1650 
1651 	/* Last segment */
1652 	/*
1653 	 * Set FIN and/or PSH flags if present only in the last packet.
1654 	 * The ip_len could be different from prior packets.
1655 	 */
1656 	bcopy(cmp->b_rptr, nmp->b_rptr, hdrlen);
1657 	nmp->b_wptr = nmp->b_rptr + hdrlen;
1658 	niph = (struct ip *)(nmp->b_rptr + ehlen);
1659 	niph->ip_id = htons(++ip_id);
1660 	niph->ip_len = htons(msgsize(nmp->b_cont) + iphlen + tcphlen);
1661 	ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1662 	tcp_seq += mss;
1663 #ifdef __sparc
1664 	tcp_seq_tmp = htonl(tcp_seq);
1665 	bcopy(&tcp_seq_tmp, (char *)&ntcph->th_seq, 4);
1666 #else
1667 	ntcph->th_seq = htonl(tcp_seq);
1668 #endif
1669 	ntcph->th_flags = (otcph->th_flags & ~TH_URG);
1670 
1671 	DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1672 	DB_CKSUMSTART(nmp) = start_offset;
1673 	DB_CKSUMSTUFF(nmp) = stuff_offset;
1674 
1675 	/* calculate IP checksum and TCP pseudo header checksum */
1676 	niph->ip_sum = 0;
1677 	niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1678 
1679 	l4_len = ntohs(niph->ip_len) - iphlen;
1680 	sum = htons(l4_len) + l4cksum;
1681 	sum = (sum & 0xFFFF) + (sum >> 16);
1682 	ntcph->th_sum = (sum & 0xffff);
1683 
1684 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1685 	    "==> nxge_do_softlso: last next "
1686 	    "niph->ip_sum 0x%x "
1687 	    "ntcph->th_sum 0x%x sum 0x%x "
1688 	    "dump last ip->tcp %s "
1689 	    "cmp $%p mp $%p (mp_chain $%p) pktlen %d (0x%x) "
1690 	    "l4_len %d (0x%x) "
1691 	    "MBLKL(mp): %d, MBLKL(datamp): %d ",
1692 	    niph->ip_sum,
1693 	    ntcph->th_sum, sum,
1694 	    nxge_dump_packet((char *)niph, 52),
1695 	    cmp, nmp, mp_chain,
1696 	    pktlen, pktlen,
1697 	    l4_len,
1698 	    l4_len,
1699 	    (int)MBLKL(mp), (int)MBLKL(datamp)));
1700 
1701 cleanup_allocated_msgs:
1702 	if (do_cleanup) {
1703 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1704 		    "==> nxge_do_softlso: "
1705 		    "Failed allocating messages, "
1706 		    "have to clean up and fail!"));
1707 		while (mp_chain != NULL) {
1708 			nmp = mp_chain;
1709 			mp_chain = mp_chain->b_next;
1710 			freemsg(nmp);
1711 		}
1712 	}
1713 	/*
1714 	 * We're done here, so just free the original message and return the
1715 	 * new message chain, that could be NULL if failed, back to the caller.
1716 	 */
1717 	freemsg(mp);
1718 
1719 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1720 	    "<== nxge_do_softlso:mp_chain $%p", mp_chain));
1721 	return (mp_chain);
1722 }
1723 
1724 /*
1725  * Will be called before NIC driver do further operation on the message.
1726  * The input message may include LSO information, if so, go to softlso logic
1727  * to eliminate the oversized LSO packet for the incapable underlying h/w.
1728  * The return could be the same non-LSO message or a message chain for LSO case.
1729  *
1730  * The driver needs to call this function per packet and process the whole chain
1731  * if applied.
1732  */
1733 static mblk_t *
1734 nxge_lso_eliminate(mblk_t *mp)
1735 {
1736 	uint32_t lsoflags;
1737 	uint32_t mss;
1738 
1739 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1740 	    "==>nxge_lso_eliminate:"));
1741 	nxge_lso_info_get(mp, &mss, &lsoflags);
1742 
1743 	if (lsoflags & HW_LSO) {
1744 		mblk_t *nmp;
1745 
1746 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1747 		    "==>nxge_lso_eliminate:"
1748 		    "HW_LSO:mss %d mp $%p",
1749 		    mss, mp));
1750 		if ((nmp = nxge_do_softlso(mp, mss)) != NULL) {
1751 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1752 			    "<== nxge_lso_eliminate: "
1753 			    "LSO: nmp not NULL nmp $%p mss %d mp $%p",
1754 			    nmp, mss, mp));
1755 			return (nmp);
1756 		} else {
1757 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1758 			    "<== nxge_lso_eliminate_ "
1759 			    "LSO: failed nmp NULL nmp $%p mss %d mp $%p",
1760 			    nmp, mss, mp));
1761 			return (NULL);
1762 		}
1763 	}
1764 
1765 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1766 	    "<== nxge_lso_eliminate"));
1767 	return (mp);
1768 }
1769 
1770 static uint32_t
1771 nxge_csgen(uint16_t *adr, int len)
1772 {
1773 	int		i, odd;
1774 	uint32_t	sum = 0;
1775 	uint32_t	c = 0;
1776 
1777 	odd = len % 2;
1778 	for (i = 0; i < (len / 2); i++) {
1779 		sum += (adr[i] & 0xffff);
1780 	}
1781 	if (odd) {
1782 		sum += adr[len / 2] & 0xff00;
1783 	}
1784 	while ((c = ((sum & 0xffff0000) >> 16)) != 0) {
1785 		sum &= 0xffff;
1786 		sum += c;
1787 	}
1788 	return (~sum & 0xffff);
1789 }
1790