xref: /illumos-gate/usr/src/uts/common/io/hxge/hxge_send.c (revision 9a5d73e03cd3312ddb571a748c40a63c58bd66e5)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <hxge_impl.h>
27 
28 extern uint32_t hxge_reclaim_pending;
29 extern uint32_t hxge_bcopy_thresh;
30 extern uint32_t hxge_dvma_thresh;
31 extern uint32_t hxge_dma_stream_thresh;
32 extern uint32_t	hxge_tx_minfree;
33 extern uint32_t	hxge_tx_intr_thres;
34 extern uint32_t	hxge_tx_max_gathers;
35 extern uint32_t	hxge_tx_tiny_pack;
36 extern uint32_t	hxge_tx_use_bcopy;
37 
38 static int hxge_start(p_hxge_t hxgep, p_tx_ring_t tx_ring_p, p_mblk_t mp);
39 
40 void
41 hxge_tx_ring_task(void *arg)
42 {
43 	p_tx_ring_t	ring = (p_tx_ring_t)arg;
44 
45 	MUTEX_ENTER(&ring->lock);
46 	(void) hxge_txdma_reclaim(ring->hxgep, ring, 0);
47 	MUTEX_EXIT(&ring->lock);
48 
49 	mac_tx_ring_update(ring->hxgep->mach, ring->ring_handle);
50 }
51 
52 static void
53 hxge_tx_ring_dispatch(p_tx_ring_t ring)
54 {
55 	/*
56 	 * Kick the ring task to reclaim some buffers.
57 	 */
58 	(void) ddi_taskq_dispatch(ring->taskq,
59 	    hxge_tx_ring_task, (void *)ring, DDI_SLEEP);
60 }
61 
62 mblk_t *
63 hxge_tx_ring_send(void *arg, mblk_t *mp)
64 {
65 	p_hxge_ring_handle_t    rhp = (p_hxge_ring_handle_t)arg;
66 	p_hxge_t		hxgep;
67 	p_tx_ring_t		tx_ring_p;
68 	int			status;
69 
70 	ASSERT(rhp != NULL);
71 	ASSERT((rhp->index >= 0) && (rhp->index < HXGE_MAX_TDCS));
72 
73 	hxgep = rhp->hxgep;
74 	tx_ring_p = hxgep->tx_rings->rings[rhp->index];
75 	ASSERT(hxgep == tx_ring_p->hxgep);
76 
77 	status = hxge_start(hxgep, tx_ring_p, mp);
78 	if (status != 0) {
79 		hxge_tx_ring_dispatch(tx_ring_p);
80 		return (mp);
81 	}
82 
83 	return ((mblk_t *)NULL);
84 }
85 
86 static int
87 hxge_start(p_hxge_t hxgep, p_tx_ring_t tx_ring_p, p_mblk_t mp)
88 {
89 	int 			status = 0;
90 	p_tx_desc_t 		tx_desc_ring_vp;
91 	hpi_handle_t		hpi_desc_handle;
92 	hxge_os_dma_handle_t 	tx_desc_dma_handle;
93 	p_tx_desc_t 		tx_desc_p;
94 	p_tx_msg_t 		tx_msg_ring;
95 	p_tx_msg_t 		tx_msg_p;
96 	tx_desc_t		tx_desc, *tmp_desc_p;
97 	tx_desc_t		sop_tx_desc, *sop_tx_desc_p;
98 	p_tx_pkt_header_t	hdrp;
99 	p_tx_pkt_hdr_all_t	pkthdrp;
100 	uint8_t			npads = 0;
101 	uint64_t 		dma_ioaddr;
102 	uint32_t		dma_flags;
103 	int			last_bidx;
104 	uint8_t 		*b_rptr;
105 	caddr_t 		kaddr;
106 	uint32_t		nmblks;
107 	uint32_t		ngathers;
108 	uint32_t		clen;
109 	int 			len;
110 	uint32_t		pkt_len, pack_len, min_len;
111 	uint32_t		bcopy_thresh;
112 	int 			i, cur_index, sop_index;
113 	uint16_t		tail_index;
114 	boolean_t		tail_wrap = B_FALSE;
115 	hxge_dma_common_t	desc_area;
116 	hxge_os_dma_handle_t 	dma_handle;
117 	ddi_dma_cookie_t 	dma_cookie;
118 	hpi_handle_t		hpi_handle;
119 	p_mblk_t 		nmp;
120 	p_mblk_t		t_mp;
121 	uint32_t 		ncookies;
122 	boolean_t 		good_packet;
123 	boolean_t 		mark_mode = B_FALSE;
124 	p_hxge_stats_t 		statsp;
125 	p_hxge_tx_ring_stats_t	tdc_stats;
126 	t_uscalar_t 		start_offset = 0;
127 	t_uscalar_t 		stuff_offset = 0;
128 	t_uscalar_t 		end_offset = 0;
129 	t_uscalar_t 		value = 0;
130 	t_uscalar_t 		cksum_flags = 0;
131 	boolean_t		cksum_on = B_FALSE;
132 	uint32_t		boff = 0;
133 	uint64_t		tot_xfer_len = 0, tmp_len = 0;
134 	boolean_t		header_set = B_FALSE;
135 	tdc_tdr_kick_t		kick;
136 	uint32_t		offset;
137 #ifdef HXGE_DEBUG
138 	p_tx_desc_t 		tx_desc_ring_pp;
139 	p_tx_desc_t 		tx_desc_pp;
140 	tx_desc_t		*save_desc_p;
141 	int			dump_len;
142 	int			sad_len;
143 	uint64_t		sad;
144 	int			xfer_len;
145 	uint32_t		msgsize;
146 #endif
147 
148 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
149 	    "==> hxge_start: tx dma channel %d", tx_ring_p->tdc));
150 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
151 	    "==> hxge_start: Starting tdc %d desc pending %d",
152 	    tx_ring_p->tdc, tx_ring_p->descs_pending));
153 
154 	statsp = hxgep->statsp;
155 
156 	if (hxgep->statsp->port_stats.lb_mode == hxge_lb_normal) {
157 		if (!statsp->mac_stats.link_up) {
158 			freemsg(mp);
159 			HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: "
160 			    "link not up or LB mode"));
161 			goto hxge_start_fail1;
162 		}
163 	}
164 
165 	hcksum_retrieve(mp, NULL, NULL, &start_offset,
166 	    &stuff_offset, &end_offset, &value, &cksum_flags);
167 	if (!HXGE_IS_VLAN_PACKET(mp->b_rptr)) {
168 		start_offset += sizeof (ether_header_t);
169 		stuff_offset += sizeof (ether_header_t);
170 	} else {
171 		start_offset += sizeof (struct ether_vlan_header);
172 		stuff_offset += sizeof (struct ether_vlan_header);
173 	}
174 
175 	if (cksum_flags & HCK_PARTIALCKSUM) {
176 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
177 		    "==> hxge_start: mp $%p len %d "
178 		    "cksum_flags 0x%x (partial checksum) ",
179 		    mp, MBLKL(mp), cksum_flags));
180 		cksum_on = B_TRUE;
181 	}
182 
183 	MUTEX_ENTER(&tx_ring_p->lock);
184 start_again:
185 	ngathers = 0;
186 	sop_index = tx_ring_p->wr_index;
187 #ifdef	HXGE_DEBUG
188 	if (tx_ring_p->descs_pending) {
189 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
190 		    "==> hxge_start: desc pending %d ",
191 		    tx_ring_p->descs_pending));
192 	}
193 
194 	dump_len = (int)(MBLKL(mp));
195 	dump_len = (dump_len > 128) ? 128: dump_len;
196 
197 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
198 	    "==> hxge_start: tdc %d: dumping ...: b_rptr $%p "
199 	    "(Before header reserve: ORIGINAL LEN %d)",
200 	    tx_ring_p->tdc, mp->b_rptr, dump_len));
201 
202 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
203 	    "==> hxge_start: dump packets (IP ORIGINAL b_rptr $%p): %s",
204 	    mp->b_rptr, hxge_dump_packet((char *)mp->b_rptr, dump_len)));
205 #endif
206 
207 	tdc_stats = tx_ring_p->tdc_stats;
208 	mark_mode = (tx_ring_p->descs_pending &&
209 	    ((tx_ring_p->tx_ring_size - tx_ring_p->descs_pending) <
210 	    hxge_tx_minfree));
211 
212 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
213 	    "TX Descriptor ring is channel %d mark mode %d",
214 	    tx_ring_p->tdc, mark_mode));
215 
216 	if (!hxge_txdma_reclaim(hxgep, tx_ring_p, hxge_tx_minfree)) {
217 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
218 		    "TX Descriptor ring is full: channel %d", tx_ring_p->tdc));
219 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
220 		    "TX Descriptor ring is full: channel %d", tx_ring_p->tdc));
221 		cas32((uint32_t *)&tx_ring_p->queueing, 0, 1);
222 		tdc_stats->tx_no_desc++;
223 		MUTEX_EXIT(&tx_ring_p->lock);
224 		status = 1;
225 		goto hxge_start_fail1;
226 	}
227 
228 	nmp = mp;
229 	i = sop_index = tx_ring_p->wr_index;
230 	nmblks = 0;
231 	ngathers = 0;
232 	pkt_len = 0;
233 	pack_len = 0;
234 	clen = 0;
235 	last_bidx = -1;
236 	good_packet = B_TRUE;
237 
238 	desc_area = tx_ring_p->tdc_desc;
239 	hpi_handle = desc_area.hpi_handle;
240 	hpi_desc_handle.regh = (hxge_os_acc_handle_t)
241 	    DMA_COMMON_ACC_HANDLE(desc_area);
242 	hpi_desc_handle.hxgep = hxgep;
243 	tx_desc_ring_vp = (p_tx_desc_t)DMA_COMMON_VPTR(desc_area);
244 #ifdef	HXGE_DEBUG
245 #if defined(__i386)
246 	tx_desc_ring_pp = (p_tx_desc_t)(uint32_t)DMA_COMMON_IOADDR(desc_area);
247 #else
248 	tx_desc_ring_pp = (p_tx_desc_t)DMA_COMMON_IOADDR(desc_area);
249 #endif
250 #endif
251 	tx_desc_dma_handle = (hxge_os_dma_handle_t)DMA_COMMON_HANDLE(desc_area);
252 	tx_msg_ring = tx_ring_p->tx_msg_ring;
253 
254 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: wr_index %d i %d",
255 	    sop_index, i));
256 
257 #ifdef	HXGE_DEBUG
258 	msgsize = msgdsize(nmp);
259 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
260 	    "==> hxge_start(1): wr_index %d i %d msgdsize %d",
261 	    sop_index, i, msgsize));
262 #endif
263 	/*
264 	 * The first 16 bytes of the premapped buffer are reserved
265 	 * for header. No padding will be used.
266 	 */
267 	pkt_len = pack_len = boff = TX_PKT_HEADER_SIZE;
268 	if (hxge_tx_use_bcopy) {
269 		bcopy_thresh = (hxge_bcopy_thresh - TX_PKT_HEADER_SIZE);
270 	} else {
271 		bcopy_thresh = (TX_BCOPY_SIZE - TX_PKT_HEADER_SIZE);
272 	}
273 	while (nmp) {
274 		good_packet = B_TRUE;
275 		b_rptr = nmp->b_rptr;
276 		len = MBLKL(nmp);
277 		if (len <= 0) {
278 			nmp = nmp->b_cont;
279 			continue;
280 		}
281 		nmblks++;
282 
283 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start(1): nmblks %d "
284 		    "len %d pkt_len %d pack_len %d",
285 		    nmblks, len, pkt_len, pack_len));
286 		/*
287 		 * Hardware limits the transfer length to 4K.
288 		 * If len is more than 4K, we need to break
289 		 * nmp into two chunks: Make first chunk smaller
290 		 * than 4K. The second chunk will be broken into
291 		 * less than 4K (if needed) during the next pass.
292 		 */
293 		if (len > (TX_MAX_TRANSFER_LENGTH - TX_PKT_HEADER_SIZE)) {
294 			if ((t_mp = dupb(nmp)) != NULL) {
295 				nmp->b_wptr = nmp->b_rptr +
296 				    (TX_MAX_TRANSFER_LENGTH -
297 				    TX_PKT_HEADER_SIZE);
298 				t_mp->b_rptr = nmp->b_wptr;
299 				t_mp->b_cont = nmp->b_cont;
300 				nmp->b_cont = t_mp;
301 				len = MBLKL(nmp);
302 			} else {
303 				good_packet = B_FALSE;
304 				goto hxge_start_fail2;
305 			}
306 		}
307 		tx_desc.value = 0;
308 		tx_desc_p = &tx_desc_ring_vp[i];
309 #ifdef	HXGE_DEBUG
310 		tx_desc_pp = &tx_desc_ring_pp[i];
311 #endif
312 		tx_msg_p = &tx_msg_ring[i];
313 #if defined(__i386)
314 		hpi_desc_handle.regp = (uint32_t)tx_desc_p;
315 #else
316 		hpi_desc_handle.regp = (uint64_t)tx_desc_p;
317 #endif
318 		if (!header_set &&
319 		    ((!hxge_tx_use_bcopy && (len > TX_BCOPY_SIZE)) ||
320 		    (len >= bcopy_thresh))) {
321 			header_set = B_TRUE;
322 			bcopy_thresh += TX_PKT_HEADER_SIZE;
323 			boff = 0;
324 			pack_len = 0;
325 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
326 			hdrp = (p_tx_pkt_header_t)kaddr;
327 			clen = pkt_len;
328 			dma_handle = tx_msg_p->buf_dma_handle;
329 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
330 			offset = tx_msg_p->offset_index * hxge_bcopy_thresh;
331 			(void) ddi_dma_sync(dma_handle,
332 			    offset, hxge_bcopy_thresh, DDI_DMA_SYNC_FORDEV);
333 
334 			tx_msg_p->flags.dma_type = USE_BCOPY;
335 			goto hxge_start_control_header_only;
336 		}
337 
338 		pkt_len += len;
339 		pack_len += len;
340 
341 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
342 		    "==> hxge_start(3): desc entry %d DESC IOADDR $%p "
343 		    "desc_vp $%p tx_desc_p $%p desc_pp $%p tx_desc_pp $%p "
344 		    "len %d pkt_len %d pack_len %d",
345 		    i,
346 		    DMA_COMMON_IOADDR(desc_area),
347 		    tx_desc_ring_vp, tx_desc_p,
348 		    tx_desc_ring_pp, tx_desc_pp,
349 		    len, pkt_len, pack_len));
350 
351 		if (len < bcopy_thresh) {
352 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
353 			    "==> hxge_start(4): USE BCOPY: "));
354 			if (hxge_tx_tiny_pack) {
355 				uint32_t blst = TXDMA_DESC_NEXT_INDEX(i, -1,
356 				    tx_ring_p->tx_wrap_mask);
357 				HXGE_DEBUG_MSG((hxgep, TX_CTL,
358 				    "==> hxge_start(5): pack"));
359 				if ((pack_len <= bcopy_thresh) &&
360 				    (last_bidx == blst)) {
361 					HXGE_DEBUG_MSG((hxgep, TX_CTL,
362 					    "==> hxge_start: pack(6) "
363 					    "(pkt_len %d pack_len %d)",
364 					    pkt_len, pack_len));
365 					i = blst;
366 					tx_desc_p = &tx_desc_ring_vp[i];
367 #ifdef	HXGE_DEBUG
368 					tx_desc_pp = &tx_desc_ring_pp[i];
369 #endif
370 					tx_msg_p = &tx_msg_ring[i];
371 					boff = pack_len - len;
372 					ngathers--;
373 				} else if (pack_len > bcopy_thresh &&
374 				    header_set) {
375 					pack_len = len;
376 					boff = 0;
377 					bcopy_thresh = hxge_bcopy_thresh;
378 					HXGE_DEBUG_MSG((hxgep, TX_CTL,
379 					    "==> hxge_start(7): > max NEW "
380 					    "bcopy thresh %d "
381 					    "pkt_len %d pack_len %d(next)",
382 					    bcopy_thresh, pkt_len, pack_len));
383 				}
384 				last_bidx = i;
385 			}
386 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
387 			if ((boff == TX_PKT_HEADER_SIZE) && (nmblks == 1)) {
388 				hdrp = (p_tx_pkt_header_t)kaddr;
389 				header_set = B_TRUE;
390 				HXGE_DEBUG_MSG((hxgep, TX_CTL,
391 				    "==> hxge_start(7_x2): "
392 				    "pkt_len %d pack_len %d (new hdrp $%p)",
393 				    pkt_len, pack_len, hdrp));
394 			}
395 			tx_msg_p->flags.dma_type = USE_BCOPY;
396 			kaddr += boff;
397 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
398 			    "==> hxge_start(8): USE BCOPY: before bcopy "
399 			    "DESC IOADDR $%p entry %d bcopy packets %d "
400 			    "bcopy kaddr $%p bcopy ioaddr (SAD) $%p "
401 			    "bcopy clen %d bcopy boff %d",
402 			    DMA_COMMON_IOADDR(desc_area), i,
403 			    tdc_stats->tx_hdr_pkts, kaddr, dma_ioaddr,
404 			    clen, boff));
405 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
406 			    "==> hxge_start: 1USE BCOPY: "));
407 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
408 			    "==> hxge_start: 2USE BCOPY: "));
409 			HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: "
410 			    "last USE BCOPY: copy from b_rptr $%p "
411 			    "to KADDR $%p (len %d offset %d",
412 			    b_rptr, kaddr, len, boff));
413 			bcopy(b_rptr, kaddr, len);
414 #ifdef	HXGE_DEBUG
415 			dump_len = (len > 128) ? 128: len;
416 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
417 			    "==> hxge_start: dump packets "
418 			    "(After BCOPY len %d)"
419 			    "(b_rptr $%p): %s", len, nmp->b_rptr,
420 			    hxge_dump_packet((char *)nmp->b_rptr,
421 			    dump_len)));
422 #endif
423 			dma_handle = tx_msg_p->buf_dma_handle;
424 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
425 			offset = tx_msg_p->offset_index * hxge_bcopy_thresh;
426 			(void) ddi_dma_sync(dma_handle,
427 			    offset, hxge_bcopy_thresh, DDI_DMA_SYNC_FORDEV);
428 			clen = len + boff;
429 			tdc_stats->tx_hdr_pkts++;
430 			HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start(9): "
431 			    "USE BCOPY: DESC IOADDR $%p entry %d "
432 			    "bcopy packets %d bcopy kaddr $%p "
433 			    "bcopy ioaddr (SAD) $%p bcopy clen %d "
434 			    "bcopy boff %d",
435 			    DMA_COMMON_IOADDR(desc_area), i,
436 			    tdc_stats->tx_hdr_pkts, kaddr, dma_ioaddr,
437 			    clen, boff));
438 		} else {
439 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
440 			    "==> hxge_start(12): USE DVMA: len %d", len));
441 			tx_msg_p->flags.dma_type = USE_DMA;
442 			dma_flags = DDI_DMA_WRITE;
443 			if (len < hxge_dma_stream_thresh) {
444 				dma_flags |= DDI_DMA_CONSISTENT;
445 			} else {
446 				dma_flags |= DDI_DMA_STREAMING;
447 			}
448 
449 			dma_handle = tx_msg_p->dma_handle;
450 			status = ddi_dma_addr_bind_handle(dma_handle, NULL,
451 			    (caddr_t)b_rptr, len, dma_flags,
452 			    DDI_DMA_DONTWAIT, NULL,
453 			    &dma_cookie, &ncookies);
454 			if (status == DDI_DMA_MAPPED) {
455 				dma_ioaddr = dma_cookie.dmac_laddress;
456 				len = (int)dma_cookie.dmac_size;
457 				clen = (uint32_t)dma_cookie.dmac_size;
458 				HXGE_DEBUG_MSG((hxgep, TX_CTL,
459 				    "==> hxge_start(12_1): "
460 				    "USE DVMA: len %d clen %d ngathers %d",
461 				    len, clen, ngathers));
462 #if defined(__i386)
463 				hpi_desc_handle.regp = (uint32_t)tx_desc_p;
464 #else
465 				hpi_desc_handle.regp = (uint64_t)tx_desc_p;
466 #endif
467 				while (ncookies > 1) {
468 					ngathers++;
469 					/*
470 					 * this is the fix for multiple
471 					 * cookies, which are basically
472 					 * a descriptor entry, we don't set
473 					 * SOP bit as well as related fields
474 					 */
475 
476 					(void) hpi_txdma_desc_gather_set(
477 					    hpi_desc_handle, &tx_desc,
478 					    (ngathers -1), mark_mode,
479 					    ngathers, dma_ioaddr, clen);
480 					tx_msg_p->tx_msg_size = clen;
481 					HXGE_DEBUG_MSG((hxgep, TX_CTL,
482 					    "==> hxge_start:  DMA "
483 					    "ncookie %d ngathers %d "
484 					    "dma_ioaddr $%p len %d"
485 					    "desc $%p descp $%p (%d)",
486 					    ncookies, ngathers,
487 					    dma_ioaddr, clen,
488 					    *tx_desc_p, tx_desc_p, i));
489 
490 					ddi_dma_nextcookie(dma_handle,
491 					    &dma_cookie);
492 					dma_ioaddr = dma_cookie.dmac_laddress;
493 
494 					len = (int)dma_cookie.dmac_size;
495 					clen = (uint32_t)dma_cookie.dmac_size;
496 					HXGE_DEBUG_MSG((hxgep, TX_CTL,
497 					    "==> hxge_start(12_2): "
498 					    "USE DVMA: len %d clen %d ",
499 					    len, clen));
500 
501 					i = TXDMA_DESC_NEXT_INDEX(i, 1,
502 					    tx_ring_p->tx_wrap_mask);
503 					tx_desc_p = &tx_desc_ring_vp[i];
504 
505 					hpi_desc_handle.regp =
506 #if defined(__i386)
507 					    (uint32_t)tx_desc_p;
508 #else
509 						(uint64_t)tx_desc_p;
510 #endif
511 					tx_msg_p = &tx_msg_ring[i];
512 					tx_msg_p->flags.dma_type = USE_NONE;
513 					tx_desc.value = 0;
514 					ncookies--;
515 				}
516 				tdc_stats->tx_ddi_pkts++;
517 				HXGE_DEBUG_MSG((hxgep, TX_CTL,
518 				    "==> hxge_start: DMA: ddi packets %d",
519 				    tdc_stats->tx_ddi_pkts));
520 			} else {
521 				HXGE_ERROR_MSG((hxgep, HXGE_ERR_CTL,
522 				    "dma mapping failed for %d "
523 				    "bytes addr $%p flags %x (%d)",
524 				    len, b_rptr, status, status));
525 				good_packet = B_FALSE;
526 				tdc_stats->tx_dma_bind_fail++;
527 				tx_msg_p->flags.dma_type = USE_NONE;
528 				goto hxge_start_fail2;
529 			}
530 		} /* ddi dvma */
531 
532 		nmp = nmp->b_cont;
533 hxge_start_control_header_only:
534 #if defined(__i386)
535 		hpi_desc_handle.regp = (uint32_t)tx_desc_p;
536 #else
537 		hpi_desc_handle.regp = (uint64_t)tx_desc_p;
538 #endif
539 		ngathers++;
540 
541 		if (ngathers == 1) {
542 #ifdef	HXGE_DEBUG
543 			save_desc_p = &sop_tx_desc;
544 #endif
545 			sop_tx_desc_p = &sop_tx_desc;
546 			sop_tx_desc_p->value = 0;
547 			sop_tx_desc_p->bits.tr_len = clen;
548 			sop_tx_desc_p->bits.sad = dma_ioaddr >> 32;
549 			sop_tx_desc_p->bits.sad_l = dma_ioaddr & 0xffffffff;
550 		} else {
551 #ifdef	HXGE_DEBUG
552 			save_desc_p = &tx_desc;
553 #endif
554 			tmp_desc_p = &tx_desc;
555 			tmp_desc_p->value = 0;
556 			tmp_desc_p->bits.tr_len = clen;
557 			tmp_desc_p->bits.sad = dma_ioaddr >> 32;
558 			tmp_desc_p->bits.sad_l = dma_ioaddr & 0xffffffff;
559 
560 			tx_desc_p->value = tmp_desc_p->value;
561 		}
562 
563 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
564 		    "==> hxge_start(13): Desc_entry %d ngathers %d "
565 		    "desc_vp $%p tx_desc_p $%p "
566 		    "len %d clen %d pkt_len %d pack_len %d nmblks %d "
567 		    "dma_ioaddr (SAD) $%p mark %d",
568 		    i, ngathers, tx_desc_ring_vp, tx_desc_p,
569 		    len, clen, pkt_len, pack_len, nmblks,
570 		    dma_ioaddr, mark_mode));
571 
572 #ifdef HXGE_DEBUG
573 		hpi_desc_handle.hxgep = hxgep;
574 		hpi_desc_handle.function.function = 0;
575 		hpi_desc_handle.function.instance = hxgep->instance;
576 		sad = save_desc_p->bits.sad;
577 		sad = (sad << 32) | save_desc_p->bits.sad_l;
578 		xfer_len = save_desc_p->bits.tr_len;
579 
580 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "\n\t: value 0x%llx\n"
581 		    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\t"
582 		    "mark %d sop %d\n",
583 		    save_desc_p->value, sad, save_desc_p->bits.tr_len,
584 		    xfer_len, save_desc_p->bits.num_ptr,
585 		    save_desc_p->bits.mark, save_desc_p->bits.sop));
586 
587 		hpi_txdma_dump_desc_one(hpi_desc_handle, NULL, i);
588 #endif
589 
590 		tx_msg_p->tx_msg_size = clen;
591 		i = TXDMA_DESC_NEXT_INDEX(i, 1, tx_ring_p->tx_wrap_mask);
592 		if (ngathers > hxge_tx_max_gathers) {
593 			good_packet = B_FALSE;
594 			hcksum_retrieve(mp, NULL, NULL, &start_offset,
595 			    &stuff_offset, &end_offset, &value, &cksum_flags);
596 
597 			HXGE_DEBUG_MSG((NULL, TX_CTL,
598 			    "==> hxge_start(14): pull msg - "
599 			    "len %d pkt_len %d ngathers %d",
600 			    len, pkt_len, ngathers));
601 			/* Pull all message blocks from b_cont */
602 			if ((msgpullup(mp, -1)) == NULL) {
603 				goto hxge_start_fail2;
604 			}
605 			goto hxge_start_fail2;
606 		}
607 	} /* while (nmp) */
608 
609 	tx_msg_p->tx_message = mp;
610 	tx_desc_p = &tx_desc_ring_vp[sop_index];
611 #if defined(__i386)
612 	hpi_desc_handle.regp = (uint32_t)tx_desc_p;
613 #else
614 	hpi_desc_handle.regp = (uint64_t)tx_desc_p;
615 #endif
616 
617 	pkthdrp = (p_tx_pkt_hdr_all_t)hdrp;
618 	pkthdrp->reserved = 0;
619 	hdrp->value = 0;
620 	(void) hxge_fill_tx_hdr(mp, B_FALSE, cksum_on,
621 	    (pkt_len - TX_PKT_HEADER_SIZE), npads, pkthdrp);
622 
623 	/*
624 	 * Hardware header should not be counted as part of the frame
625 	 * when determining the frame size
626 	 */
627 	if ((pkt_len - TX_PKT_HEADER_SIZE) > (STD_FRAME_SIZE - ETHERFCSL)) {
628 		tdc_stats->tx_jumbo_pkts++;
629 	}
630 
631 	min_len = (hxgep->msg_min + TX_PKT_HEADER_SIZE + (npads * 2));
632 	if (pkt_len < min_len) {
633 		/* Assume we use bcopy to premapped buffers */
634 		kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
635 		HXGE_DEBUG_MSG((NULL, TX_CTL,
636 		    "==> hxge_start(14-1): < (msg_min + 16)"
637 		    "len %d pkt_len %d min_len %d bzero %d ngathers %d",
638 		    len, pkt_len, min_len, (min_len - pkt_len), ngathers));
639 		bzero((kaddr + pkt_len), (min_len - pkt_len));
640 		pkt_len = tx_msg_p->tx_msg_size = min_len;
641 
642 		sop_tx_desc_p->bits.tr_len = min_len;
643 
644 		HXGE_MEM_PIO_WRITE64(hpi_desc_handle, sop_tx_desc_p->value);
645 		tx_desc_p->value = sop_tx_desc_p->value;
646 
647 		HXGE_DEBUG_MSG((NULL, TX_CTL,
648 		    "==> hxge_start(14-2): < msg_min - "
649 		    "len %d pkt_len %d min_len %d ngathers %d",
650 		    len, pkt_len, min_len, ngathers));
651 	}
652 
653 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: cksum_flags 0x%x ",
654 	    cksum_flags));
655 	if (cksum_flags & HCK_PARTIALCKSUM) {
656 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
657 		    "==> hxge_start: cksum_flags 0x%x (partial checksum) ",
658 		    cksum_flags));
659 		cksum_on = B_TRUE;
660 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
661 		    "==> hxge_start: from IP cksum_flags 0x%x "
662 		    "(partial checksum) "
663 		    "start_offset %d stuff_offset %d",
664 		    cksum_flags, start_offset, stuff_offset));
665 		tmp_len = (uint64_t)(start_offset >> 1);
666 		hdrp->value |= (tmp_len << TX_PKT_HEADER_L4START_SHIFT);
667 		tmp_len = (uint64_t)(stuff_offset >> 1);
668 		hdrp->value |= (tmp_len << TX_PKT_HEADER_L4STUFF_SHIFT);
669 
670 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
671 		    "==> hxge_start: from IP cksum_flags 0x%x "
672 		    "(partial checksum) "
673 		    "after SHIFT start_offset %d stuff_offset %d",
674 		    cksum_flags, start_offset, stuff_offset));
675 	}
676 
677 	/*
678 	 * pkt_len already includes 16 + paddings!!
679 	 * Update the control header length
680 	 */
681 
682 	/*
683 	 * Note that Hydra is different from Neptune where
684 	 * tot_xfer_len = (pkt_len - TX_PKT_HEADER_SIZE);
685 	 */
686 	tot_xfer_len = pkt_len;
687 	tmp_len = hdrp->value |
688 	    (tot_xfer_len << TX_PKT_HEADER_TOT_XFER_LEN_SHIFT);
689 
690 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
691 	    "==> hxge_start(15_x1): setting SOP "
692 	    "tot_xfer_len 0x%llx (%d) pkt_len %d tmp_len "
693 	    "0x%llx hdrp->value 0x%llx",
694 	    tot_xfer_len, tot_xfer_len, pkt_len, tmp_len, hdrp->value));
695 #if defined(_BIG_ENDIAN)
696 	hdrp->value = ddi_swap64(tmp_len);
697 #else
698 	hdrp->value = tmp_len;
699 #endif
700 	HXGE_DEBUG_MSG((hxgep,
701 	    TX_CTL, "==> hxge_start(15_x2): setting SOP "
702 	    "after SWAP: tot_xfer_len 0x%llx pkt_len %d "
703 	    "tmp_len 0x%llx hdrp->value 0x%llx",
704 	    tot_xfer_len, pkt_len, tmp_len, hdrp->value));
705 
706 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start(15): setting SOP "
707 	    "wr_index %d tot_xfer_len (%d) pkt_len %d npads %d",
708 	    sop_index, tot_xfer_len, pkt_len, npads));
709 
710 	sop_tx_desc_p->bits.sop = 1;
711 	sop_tx_desc_p->bits.mark = mark_mode;
712 	sop_tx_desc_p->bits.num_ptr = ngathers;
713 
714 	if (mark_mode)
715 		tdc_stats->tx_marks++;
716 
717 	HXGE_MEM_PIO_WRITE64(hpi_desc_handle, sop_tx_desc_p->value);
718 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start(16): set SOP done"));
719 
720 #ifdef HXGE_DEBUG
721 	hpi_desc_handle.hxgep = hxgep;
722 	hpi_desc_handle.function.function = 0;
723 	hpi_desc_handle.function.instance = hxgep->instance;
724 
725 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "\n\t: value 0x%llx\n"
726 	    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\tmark %d sop %d\n",
727 	    save_desc_p->value, sad, save_desc_p->bits.tr_len,
728 	    xfer_len, save_desc_p->bits.num_ptr, save_desc_p->bits.mark,
729 	    save_desc_p->bits.sop));
730 	(void) hpi_txdma_dump_desc_one(hpi_desc_handle, NULL, sop_index);
731 
732 	dump_len = (pkt_len > 128) ? 128: pkt_len;
733 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
734 	    "==> hxge_start: dump packets(17) (after sop set, len "
735 	    " (len/dump_len/pkt_len/tot_xfer_len) %d/%d/%d/%d):\n"
736 	    "ptr $%p: %s", len, dump_len, pkt_len, tot_xfer_len,
737 	    (char *)hdrp, hxge_dump_packet((char *)hdrp, dump_len)));
738 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
739 	    "==> hxge_start(18): TX desc sync: sop_index %d", sop_index));
740 #endif
741 
742 	if ((ngathers == 1) || tx_ring_p->wr_index < i) {
743 		(void) ddi_dma_sync(tx_desc_dma_handle,
744 		    sop_index * sizeof (tx_desc_t),
745 		    ngathers * sizeof (tx_desc_t), DDI_DMA_SYNC_FORDEV);
746 
747 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "hxge_start(19): sync 1 "
748 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
749 		    "pkt_len %d ngathers %d sop_index %d\n",
750 		    stuff_offset, start_offset,
751 		    pkt_len, ngathers, sop_index));
752 	} else { /* more than one descriptor and wrap around */
753 		uint32_t nsdescs = tx_ring_p->tx_ring_size - sop_index;
754 		(void) ddi_dma_sync(tx_desc_dma_handle,
755 		    sop_index * sizeof (tx_desc_t),
756 		    nsdescs * sizeof (tx_desc_t), DDI_DMA_SYNC_FORDEV);
757 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "hxge_start(20): sync 1 "
758 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
759 		    "pkt_len %d ngathers %d sop_index %d\n",
760 		    stuff_offset, start_offset, pkt_len, ngathers, sop_index));
761 
762 		(void) ddi_dma_sync(tx_desc_dma_handle, 0,
763 		    (ngathers - nsdescs) * sizeof (tx_desc_t),
764 		    DDI_DMA_SYNC_FORDEV);
765 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "hxge_start(21): sync 2 "
766 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
767 		    "pkt_len %d ngathers %d sop_index %d\n",
768 		    stuff_offset, start_offset,
769 		    pkt_len, ngathers, sop_index));
770 	}
771 
772 	tail_index = tx_ring_p->wr_index;
773 	tail_wrap = tx_ring_p->wr_index_wrap;
774 
775 	tx_ring_p->wr_index = i;
776 	if (tx_ring_p->wr_index <= tail_index) {
777 		tx_ring_p->wr_index_wrap = ((tail_wrap == B_TRUE) ?
778 		    B_FALSE : B_TRUE);
779 	}
780 
781 	tx_ring_p->descs_pending += ngathers;
782 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: TX kick: "
783 	    "channel %d wr_index %d wrap %d ngathers %d desc_pend %d",
784 	    tx_ring_p->tdc, tx_ring_p->wr_index, tx_ring_p->wr_index_wrap,
785 	    ngathers, tx_ring_p->descs_pending));
786 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: TX KICKING: "));
787 
788 	kick.value = 0;
789 	kick.bits.wrap = tx_ring_p->wr_index_wrap;
790 	kick.bits.tail = (uint16_t)tx_ring_p->wr_index;
791 
792 	/* Kick start the Transmit kick register */
793 	TXDMA_REG_WRITE64(HXGE_DEV_HPI_HANDLE(hxgep),
794 	    TDC_TDR_KICK, (uint8_t)tx_ring_p->tdc, kick.value);
795 	tdc_stats->tx_starts++;
796 	MUTEX_EXIT(&tx_ring_p->lock);
797 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "<== hxge_start"));
798 	return (status);
799 
800 hxge_start_fail2:
801 	if (good_packet == B_FALSE) {
802 		cur_index = sop_index;
803 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: clean up"));
804 		for (i = 0; i < ngathers; i++) {
805 			tx_desc_p = &tx_desc_ring_vp[cur_index];
806 #if defined(__i386)
807 			hpi_handle.regp = (uint32_t)tx_desc_p;
808 #else
809 			hpi_handle.regp = (uint64_t)tx_desc_p;
810 #endif
811 			tx_msg_p = &tx_msg_ring[cur_index];
812 			(void) hpi_txdma_desc_set_zero(hpi_handle, 1);
813 			if (tx_msg_p->flags.dma_type == USE_DVMA) {
814 				HXGE_DEBUG_MSG((hxgep, TX_CTL,
815 				    "tx_desc_p = %X index = %d",
816 				    tx_desc_p, tx_ring_p->rd_index));
817 				(void) dvma_unload(tx_msg_p->dvma_handle,
818 				    0, -1);
819 				tx_msg_p->dvma_handle = NULL;
820 				if (tx_ring_p->dvma_wr_index ==
821 				    tx_ring_p->dvma_wrap_mask)
822 					tx_ring_p->dvma_wr_index = 0;
823 				else
824 					tx_ring_p->dvma_wr_index++;
825 				tx_ring_p->dvma_pending--;
826 			} else if (tx_msg_p->flags.dma_type == USE_DMA) {
827 				if (ddi_dma_unbind_handle(
828 				    tx_msg_p->dma_handle)) {
829 					cmn_err(CE_WARN, "hxge_start: "
830 					    "ddi_dma_unbind_handle failed");
831 				}
832 			}
833 			tx_msg_p->flags.dma_type = USE_NONE;
834 			cur_index = TXDMA_DESC_NEXT_INDEX(cur_index, 1,
835 			    tx_ring_p->tx_wrap_mask);
836 
837 		}
838 	}
839 
840 	MUTEX_EXIT(&tx_ring_p->lock);
841 
842 hxge_start_fail1:
843 	/* Add FMA to check the access handle hxge_hregh */
844 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "<== hxge_start"));
845 	return (status);
846 }
847