xref: /titanic_50/usr/src/uts/common/io/hxge/hxge_send.c (revision e58a33b62cd4c9a6815fd752ce58b5f389289da1)
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 /*
23  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #include <hxge_impl.h>
28 
29 extern uint32_t hxge_reclaim_pending;
30 extern uint32_t hxge_bcopy_thresh;
31 extern uint32_t hxge_dvma_thresh;
32 extern uint32_t hxge_dma_stream_thresh;
33 extern uint32_t	hxge_tx_minfree;
34 extern uint32_t	hxge_tx_intr_thres;
35 extern uint32_t	hxge_tx_max_gathers;
36 extern uint32_t	hxge_tx_tiny_pack;
37 extern uint32_t	hxge_tx_use_bcopy;
38 
39 static int hxge_start(p_hxge_t hxgep, p_tx_ring_t tx_ring_p, p_mblk_t mp);
40 
41 void
42 hxge_tx_ring_task(void *arg)
43 {
44 	p_tx_ring_t	ring = (p_tx_ring_t)arg;
45 
46 	MUTEX_ENTER(&ring->lock);
47 	(void) hxge_txdma_reclaim(ring->hxgep, ring, 0);
48 	MUTEX_EXIT(&ring->lock);
49 
50 	mac_tx_ring_update(ring->hxgep->mach, ring->ring_handle);
51 }
52 
53 static void
54 hxge_tx_ring_dispatch(p_tx_ring_t ring)
55 {
56 	/*
57 	 * Kick the ring task to reclaim some buffers.
58 	 */
59 	(void) ddi_taskq_dispatch(ring->taskq,
60 	    hxge_tx_ring_task, (void *)ring, DDI_SLEEP);
61 }
62 
63 mblk_t *
64 hxge_tx_ring_send(void *arg, mblk_t *mp)
65 {
66 	p_hxge_ring_handle_t    rhp = (p_hxge_ring_handle_t)arg;
67 	p_hxge_t		hxgep;
68 	p_tx_ring_t		tx_ring_p;
69 	int			status;
70 
71 	ASSERT(rhp != NULL);
72 	ASSERT((rhp->index >= 0) && (rhp->index < HXGE_MAX_TDCS));
73 
74 	hxgep = rhp->hxgep;
75 	tx_ring_p = hxgep->tx_rings->rings[rhp->index];
76 	ASSERT(hxgep == tx_ring_p->hxgep);
77 
78 	status = hxge_start(hxgep, tx_ring_p, mp);
79 	if (status != 0) {
80 		hxge_tx_ring_dispatch(tx_ring_p);
81 		return (mp);
82 	}
83 
84 	return ((mblk_t *)NULL);
85 }
86 
87 static int
88 hxge_start(p_hxge_t hxgep, p_tx_ring_t tx_ring_p, p_mblk_t mp)
89 {
90 	int 			dma_status, status = 0;
91 	p_tx_desc_t 		tx_desc_ring_vp;
92 	hpi_handle_t		hpi_desc_handle;
93 	hxge_os_dma_handle_t 	tx_desc_dma_handle;
94 	p_tx_desc_t 		tx_desc_p;
95 	p_tx_msg_t 		tx_msg_ring;
96 	p_tx_msg_t 		tx_msg_p;
97 	tx_desc_t		tx_desc, *tmp_desc_p;
98 	tx_desc_t		sop_tx_desc, *sop_tx_desc_p;
99 	p_tx_pkt_header_t	hdrp;
100 	p_tx_pkt_hdr_all_t	pkthdrp;
101 	uint8_t			npads = 0;
102 	uint64_t 		dma_ioaddr;
103 	uint32_t		dma_flags;
104 	int			last_bidx;
105 	uint8_t 		*b_rptr;
106 	caddr_t 		kaddr;
107 	uint32_t		nmblks;
108 	uint32_t		ngathers;
109 	uint32_t		clen;
110 	int 			len;
111 	uint32_t		pkt_len, pack_len, min_len;
112 	uint32_t		bcopy_thresh;
113 	int 			i, cur_index, sop_index;
114 	uint16_t		tail_index;
115 	boolean_t		tail_wrap = B_FALSE;
116 	hxge_dma_common_t	desc_area;
117 	hxge_os_dma_handle_t 	dma_handle;
118 	ddi_dma_cookie_t 	dma_cookie;
119 	hpi_handle_t		hpi_handle;
120 	p_mblk_t 		nmp;
121 	p_mblk_t		t_mp;
122 	uint32_t 		ncookies;
123 	boolean_t 		good_packet;
124 	boolean_t 		mark_mode = B_FALSE;
125 	p_hxge_stats_t 		statsp;
126 	p_hxge_tx_ring_stats_t	tdc_stats;
127 	t_uscalar_t 		start_offset = 0;
128 	t_uscalar_t 		stuff_offset = 0;
129 	t_uscalar_t 		end_offset = 0;
130 	t_uscalar_t 		value = 0;
131 	t_uscalar_t 		cksum_flags = 0;
132 	boolean_t		cksum_on = B_FALSE;
133 	uint32_t		boff = 0;
134 	uint64_t		tot_xfer_len = 0, tmp_len = 0;
135 	boolean_t		header_set = B_FALSE;
136 	tdc_tdr_kick_t		kick;
137 	uint32_t		offset;
138 #ifdef HXGE_DEBUG
139 	p_tx_desc_t 		tx_desc_ring_pp;
140 	p_tx_desc_t 		tx_desc_pp;
141 	tx_desc_t		*save_desc_p;
142 	int			dump_len;
143 	int			sad_len;
144 	uint64_t		sad;
145 	int			xfer_len;
146 	uint32_t		msgsize;
147 #endif
148 
149 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
150 	    "==> hxge_start: tx dma channel %d", tx_ring_p->tdc));
151 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
152 	    "==> hxge_start: Starting tdc %d desc pending %d",
153 	    tx_ring_p->tdc, tx_ring_p->descs_pending));
154 
155 	statsp = hxgep->statsp;
156 
157 	if (hxgep->statsp->port_stats.lb_mode == hxge_lb_normal) {
158 		if (!statsp->mac_stats.link_up) {
159 			freemsg(mp);
160 			HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: "
161 			    "link not up or LB mode"));
162 			goto hxge_start_fail1;
163 		}
164 	}
165 
166 	mac_hcksum_get(mp, &start_offset, &stuff_offset, &end_offset, &value,
167 	    &cksum_flags);
168 	if (!HXGE_IS_VLAN_PACKET(mp->b_rptr)) {
169 		start_offset += sizeof (ether_header_t);
170 		stuff_offset += sizeof (ether_header_t);
171 	} else {
172 		start_offset += sizeof (struct ether_vlan_header);
173 		stuff_offset += sizeof (struct ether_vlan_header);
174 	}
175 
176 	if (cksum_flags & HCK_PARTIALCKSUM) {
177 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
178 		    "==> hxge_start: mp $%p len %d "
179 		    "cksum_flags 0x%x (partial checksum) ",
180 		    mp, MBLKL(mp), cksum_flags));
181 		cksum_on = B_TRUE;
182 	}
183 
184 	MUTEX_ENTER(&tx_ring_p->lock);
185 start_again:
186 	ngathers = 0;
187 	sop_index = tx_ring_p->wr_index;
188 #ifdef	HXGE_DEBUG
189 	if (tx_ring_p->descs_pending) {
190 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
191 		    "==> hxge_start: desc pending %d ",
192 		    tx_ring_p->descs_pending));
193 	}
194 
195 	dump_len = (int)(MBLKL(mp));
196 	dump_len = (dump_len > 128) ? 128: dump_len;
197 
198 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
199 	    "==> hxge_start: tdc %d: dumping ...: b_rptr $%p "
200 	    "(Before header reserve: ORIGINAL LEN %d)",
201 	    tx_ring_p->tdc, mp->b_rptr, dump_len));
202 
203 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
204 	    "==> hxge_start: dump packets (IP ORIGINAL b_rptr $%p): %s",
205 	    mp->b_rptr, hxge_dump_packet((char *)mp->b_rptr, dump_len)));
206 #endif
207 
208 	tdc_stats = tx_ring_p->tdc_stats;
209 	mark_mode = (tx_ring_p->descs_pending &&
210 	    ((tx_ring_p->tx_ring_size - tx_ring_p->descs_pending) <
211 	    hxge_tx_minfree));
212 
213 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
214 	    "TX Descriptor ring is channel %d mark mode %d",
215 	    tx_ring_p->tdc, mark_mode));
216 
217 	if (!hxge_txdma_reclaim(hxgep, tx_ring_p, hxge_tx_minfree)) {
218 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
219 		    "TX Descriptor ring is full: channel %d", tx_ring_p->tdc));
220 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
221 		    "TX Descriptor ring is full: channel %d", tx_ring_p->tdc));
222 		(void) cas32((uint32_t *)&tx_ring_p->queueing, 0, 1);
223 		tdc_stats->tx_no_desc++;
224 		MUTEX_EXIT(&tx_ring_p->lock);
225 		status = 1;
226 		goto hxge_start_fail1;
227 	}
228 
229 	nmp = mp;
230 	i = sop_index = tx_ring_p->wr_index;
231 	nmblks = 0;
232 	ngathers = 0;
233 	pkt_len = 0;
234 	pack_len = 0;
235 	clen = 0;
236 	last_bidx = -1;
237 	good_packet = B_TRUE;
238 
239 	desc_area = tx_ring_p->tdc_desc;
240 	hpi_handle = desc_area.hpi_handle;
241 	hpi_desc_handle.regh = (hxge_os_acc_handle_t)
242 	    DMA_COMMON_ACC_HANDLE(desc_area);
243 	hpi_desc_handle.hxgep = hxgep;
244 	tx_desc_ring_vp = (p_tx_desc_t)DMA_COMMON_VPTR(desc_area);
245 #ifdef	HXGE_DEBUG
246 #if defined(__i386)
247 	tx_desc_ring_pp = (p_tx_desc_t)(uint32_t)DMA_COMMON_IOADDR(desc_area);
248 #else
249 	tx_desc_ring_pp = (p_tx_desc_t)DMA_COMMON_IOADDR(desc_area);
250 #endif
251 #endif
252 	tx_desc_dma_handle = (hxge_os_dma_handle_t)DMA_COMMON_HANDLE(desc_area);
253 	tx_msg_ring = tx_ring_p->tx_msg_ring;
254 
255 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: wr_index %d i %d",
256 	    sop_index, i));
257 
258 #ifdef	HXGE_DEBUG
259 	msgsize = msgdsize(nmp);
260 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
261 	    "==> hxge_start(1): wr_index %d i %d msgdsize %d",
262 	    sop_index, i, msgsize));
263 #endif
264 	/*
265 	 * The first 16 bytes of the premapped buffer are reserved
266 	 * for header. No padding will be used.
267 	 */
268 	pkt_len = pack_len = boff = TX_PKT_HEADER_SIZE;
269 	if (hxge_tx_use_bcopy) {
270 		bcopy_thresh = (hxge_bcopy_thresh - TX_PKT_HEADER_SIZE);
271 	} else {
272 		bcopy_thresh = (TX_BCOPY_SIZE - TX_PKT_HEADER_SIZE);
273 	}
274 	while (nmp) {
275 		good_packet = B_TRUE;
276 		b_rptr = nmp->b_rptr;
277 		len = MBLKL(nmp);
278 		if (len <= 0) {
279 			nmp = nmp->b_cont;
280 			continue;
281 		}
282 		nmblks++;
283 
284 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start(1): nmblks %d "
285 		    "len %d pkt_len %d pack_len %d",
286 		    nmblks, len, pkt_len, pack_len));
287 		/*
288 		 * Hardware limits the transfer length to 4K.
289 		 * If len is more than 4K, we need to break
290 		 * nmp into two chunks: Make first chunk smaller
291 		 * than 4K. The second chunk will be broken into
292 		 * less than 4K (if needed) during the next pass.
293 		 */
294 		if (len > (TX_MAX_TRANSFER_LENGTH - TX_PKT_HEADER_SIZE)) {
295 			if ((t_mp = dupb(nmp)) != NULL) {
296 				nmp->b_wptr = nmp->b_rptr +
297 				    (TX_MAX_TRANSFER_LENGTH -
298 				    TX_PKT_HEADER_SIZE);
299 				t_mp->b_rptr = nmp->b_wptr;
300 				t_mp->b_cont = nmp->b_cont;
301 				nmp->b_cont = t_mp;
302 				len = MBLKL(nmp);
303 			} else {
304 				good_packet = B_FALSE;
305 				goto hxge_start_fail2;
306 			}
307 		}
308 		tx_desc.value = 0;
309 		tx_desc_p = &tx_desc_ring_vp[i];
310 #ifdef	HXGE_DEBUG
311 		tx_desc_pp = &tx_desc_ring_pp[i];
312 #endif
313 		tx_msg_p = &tx_msg_ring[i];
314 #if defined(__i386)
315 		hpi_desc_handle.regp = (uint32_t)tx_desc_p;
316 #else
317 		hpi_desc_handle.regp = (uint64_t)tx_desc_p;
318 #endif
319 		if (!header_set &&
320 		    ((!hxge_tx_use_bcopy && (len > TX_BCOPY_SIZE)) ||
321 		    (len >= bcopy_thresh))) {
322 			header_set = B_TRUE;
323 			bcopy_thresh += TX_PKT_HEADER_SIZE;
324 			boff = 0;
325 			pack_len = 0;
326 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
327 			hdrp = (p_tx_pkt_header_t)kaddr;
328 			clen = pkt_len;
329 			dma_handle = tx_msg_p->buf_dma_handle;
330 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
331 			offset = tx_msg_p->offset_index * hxge_bcopy_thresh;
332 			(void) ddi_dma_sync(dma_handle,
333 			    offset, hxge_bcopy_thresh, DDI_DMA_SYNC_FORDEV);
334 
335 			tx_msg_p->flags.dma_type = USE_BCOPY;
336 			goto hxge_start_control_header_only;
337 		}
338 
339 		pkt_len += len;
340 		pack_len += len;
341 
342 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
343 		    "==> hxge_start(3): desc entry %d DESC IOADDR $%p "
344 		    "desc_vp $%p tx_desc_p $%p desc_pp $%p tx_desc_pp $%p "
345 		    "len %d pkt_len %d pack_len %d",
346 		    i,
347 		    DMA_COMMON_IOADDR(desc_area),
348 		    tx_desc_ring_vp, tx_desc_p,
349 		    tx_desc_ring_pp, tx_desc_pp,
350 		    len, pkt_len, pack_len));
351 
352 		if (len < bcopy_thresh) {
353 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
354 			    "==> hxge_start(4): USE BCOPY: "));
355 			if (hxge_tx_tiny_pack) {
356 				uint32_t blst = TXDMA_DESC_NEXT_INDEX(i, -1,
357 				    tx_ring_p->tx_wrap_mask);
358 				HXGE_DEBUG_MSG((hxgep, TX_CTL,
359 				    "==> hxge_start(5): pack"));
360 				if ((pack_len <= bcopy_thresh) &&
361 				    (last_bidx == blst)) {
362 					HXGE_DEBUG_MSG((hxgep, TX_CTL,
363 					    "==> hxge_start: pack(6) "
364 					    "(pkt_len %d pack_len %d)",
365 					    pkt_len, pack_len));
366 					i = blst;
367 					tx_desc_p = &tx_desc_ring_vp[i];
368 #ifdef	HXGE_DEBUG
369 					tx_desc_pp = &tx_desc_ring_pp[i];
370 #endif
371 					tx_msg_p = &tx_msg_ring[i];
372 					boff = pack_len - len;
373 					ngathers--;
374 				} else if (pack_len > bcopy_thresh &&
375 				    header_set) {
376 					pack_len = len;
377 					boff = 0;
378 					bcopy_thresh = hxge_bcopy_thresh;
379 					HXGE_DEBUG_MSG((hxgep, TX_CTL,
380 					    "==> hxge_start(7): > max NEW "
381 					    "bcopy thresh %d "
382 					    "pkt_len %d pack_len %d(next)",
383 					    bcopy_thresh, pkt_len, pack_len));
384 				}
385 				last_bidx = i;
386 			}
387 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
388 			if ((boff == TX_PKT_HEADER_SIZE) && (nmblks == 1)) {
389 				hdrp = (p_tx_pkt_header_t)kaddr;
390 				header_set = B_TRUE;
391 				HXGE_DEBUG_MSG((hxgep, TX_CTL,
392 				    "==> hxge_start(7_x2): "
393 				    "pkt_len %d pack_len %d (new hdrp $%p)",
394 				    pkt_len, pack_len, hdrp));
395 			}
396 			tx_msg_p->flags.dma_type = USE_BCOPY;
397 			kaddr += boff;
398 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
399 			    "==> hxge_start(8): USE BCOPY: before bcopy "
400 			    "DESC IOADDR $%p entry %d bcopy packets %d "
401 			    "bcopy kaddr $%p bcopy ioaddr (SAD) $%p "
402 			    "bcopy clen %d bcopy boff %d",
403 			    DMA_COMMON_IOADDR(desc_area), i,
404 			    tdc_stats->tx_hdr_pkts, kaddr, dma_ioaddr,
405 			    clen, boff));
406 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
407 			    "==> hxge_start: 1USE BCOPY: "));
408 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
409 			    "==> hxge_start: 2USE BCOPY: "));
410 			HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: "
411 			    "last USE BCOPY: copy from b_rptr $%p "
412 			    "to KADDR $%p (len %d offset %d",
413 			    b_rptr, kaddr, len, boff));
414 			bcopy(b_rptr, kaddr, len);
415 #ifdef	HXGE_DEBUG
416 			dump_len = (len > 128) ? 128: len;
417 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
418 			    "==> hxge_start: dump packets "
419 			    "(After BCOPY len %d)"
420 			    "(b_rptr $%p): %s", len, nmp->b_rptr,
421 			    hxge_dump_packet((char *)nmp->b_rptr,
422 			    dump_len)));
423 #endif
424 			dma_handle = tx_msg_p->buf_dma_handle;
425 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
426 			offset = tx_msg_p->offset_index * hxge_bcopy_thresh;
427 			(void) ddi_dma_sync(dma_handle,
428 			    offset, hxge_bcopy_thresh, DDI_DMA_SYNC_FORDEV);
429 			clen = len + boff;
430 			tdc_stats->tx_hdr_pkts++;
431 			HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start(9): "
432 			    "USE BCOPY: DESC IOADDR $%p entry %d "
433 			    "bcopy packets %d bcopy kaddr $%p "
434 			    "bcopy ioaddr (SAD) $%p bcopy clen %d "
435 			    "bcopy boff %d",
436 			    DMA_COMMON_IOADDR(desc_area), i,
437 			    tdc_stats->tx_hdr_pkts, kaddr, dma_ioaddr,
438 			    clen, boff));
439 		} else {
440 			HXGE_DEBUG_MSG((hxgep, TX_CTL,
441 			    "==> hxge_start(12): USE DVMA: len %d", len));
442 			tx_msg_p->flags.dma_type = USE_DMA;
443 			dma_flags = DDI_DMA_WRITE;
444 			if (len < hxge_dma_stream_thresh) {
445 				dma_flags |= DDI_DMA_CONSISTENT;
446 			} else {
447 				dma_flags |= DDI_DMA_STREAMING;
448 			}
449 
450 			dma_handle = tx_msg_p->dma_handle;
451 			dma_status = ddi_dma_addr_bind_handle(dma_handle, NULL,
452 			    (caddr_t)b_rptr, len, dma_flags,
453 			    DDI_DMA_DONTWAIT, NULL,
454 			    &dma_cookie, &ncookies);
455 			if (dma_status == DDI_DMA_MAPPED) {
456 				dma_ioaddr = dma_cookie.dmac_laddress;
457 				len = (int)dma_cookie.dmac_size;
458 				clen = (uint32_t)dma_cookie.dmac_size;
459 				HXGE_DEBUG_MSG((hxgep, TX_CTL,
460 				    "==> hxge_start(12_1): "
461 				    "USE DVMA: len %d clen %d ngathers %d",
462 				    len, clen, ngathers));
463 #if defined(__i386)
464 				hpi_desc_handle.regp = (uint32_t)tx_desc_p;
465 #else
466 				hpi_desc_handle.regp = (uint64_t)tx_desc_p;
467 #endif
468 				while (ncookies > 1) {
469 					ngathers++;
470 					/*
471 					 * this is the fix for multiple
472 					 * cookies, which are basically
473 					 * a descriptor entry, we don't set
474 					 * SOP bit as well as related fields
475 					 */
476 
477 					(void) hpi_txdma_desc_gather_set(
478 					    hpi_desc_handle, &tx_desc,
479 					    (ngathers -1), mark_mode,
480 					    ngathers, dma_ioaddr, clen);
481 					tx_msg_p->tx_msg_size = clen;
482 					HXGE_DEBUG_MSG((hxgep, TX_CTL,
483 					    "==> hxge_start:  DMA "
484 					    "ncookie %d ngathers %d "
485 					    "dma_ioaddr $%p len %d"
486 					    "desc $%p descp $%p (%d)",
487 					    ncookies, ngathers,
488 					    dma_ioaddr, clen,
489 					    *tx_desc_p, tx_desc_p, i));
490 
491 					ddi_dma_nextcookie(dma_handle,
492 					    &dma_cookie);
493 					dma_ioaddr = dma_cookie.dmac_laddress;
494 
495 					len = (int)dma_cookie.dmac_size;
496 					clen = (uint32_t)dma_cookie.dmac_size;
497 					HXGE_DEBUG_MSG((hxgep, TX_CTL,
498 					    "==> hxge_start(12_2): "
499 					    "USE DVMA: len %d clen %d ",
500 					    len, clen));
501 
502 					i = TXDMA_DESC_NEXT_INDEX(i, 1,
503 					    tx_ring_p->tx_wrap_mask);
504 					tx_desc_p = &tx_desc_ring_vp[i];
505 
506 					hpi_desc_handle.regp =
507 #if defined(__i386)
508 					    (uint32_t)tx_desc_p;
509 #else
510 						(uint64_t)tx_desc_p;
511 #endif
512 					tx_msg_p = &tx_msg_ring[i];
513 					tx_msg_p->flags.dma_type = USE_NONE;
514 					tx_desc.value = 0;
515 					ncookies--;
516 				}
517 				tdc_stats->tx_ddi_pkts++;
518 				HXGE_DEBUG_MSG((hxgep, TX_CTL,
519 				    "==> hxge_start: DMA: ddi packets %d",
520 				    tdc_stats->tx_ddi_pkts));
521 			} else {
522 				HXGE_ERROR_MSG((hxgep, HXGE_ERR_CTL,
523 				    "dma mapping failed for %d "
524 				    "bytes addr $%p flags %x (%d)",
525 				    len, b_rptr, status, status));
526 				good_packet = B_FALSE;
527 				tdc_stats->tx_dma_bind_fail++;
528 				tx_msg_p->flags.dma_type = USE_NONE;
529 				status = 1;
530 				goto hxge_start_fail2;
531 			}
532 		} /* ddi dvma */
533 
534 		nmp = nmp->b_cont;
535 hxge_start_control_header_only:
536 #if defined(__i386)
537 		hpi_desc_handle.regp = (uint32_t)tx_desc_p;
538 #else
539 		hpi_desc_handle.regp = (uint64_t)tx_desc_p;
540 #endif
541 		ngathers++;
542 
543 		if (ngathers == 1) {
544 #ifdef	HXGE_DEBUG
545 			save_desc_p = &sop_tx_desc;
546 #endif
547 			sop_tx_desc_p = &sop_tx_desc;
548 			sop_tx_desc_p->value = 0;
549 			sop_tx_desc_p->bits.tr_len = clen;
550 			sop_tx_desc_p->bits.sad = dma_ioaddr >> 32;
551 			sop_tx_desc_p->bits.sad_l = dma_ioaddr & 0xffffffff;
552 		} else {
553 #ifdef	HXGE_DEBUG
554 			save_desc_p = &tx_desc;
555 #endif
556 			tmp_desc_p = &tx_desc;
557 			tmp_desc_p->value = 0;
558 			tmp_desc_p->bits.tr_len = clen;
559 			tmp_desc_p->bits.sad = dma_ioaddr >> 32;
560 			tmp_desc_p->bits.sad_l = dma_ioaddr & 0xffffffff;
561 
562 			tx_desc_p->value = tmp_desc_p->value;
563 		}
564 
565 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
566 		    "==> hxge_start(13): Desc_entry %d ngathers %d "
567 		    "desc_vp $%p tx_desc_p $%p "
568 		    "len %d clen %d pkt_len %d pack_len %d nmblks %d "
569 		    "dma_ioaddr (SAD) $%p mark %d",
570 		    i, ngathers, tx_desc_ring_vp, tx_desc_p,
571 		    len, clen, pkt_len, pack_len, nmblks,
572 		    dma_ioaddr, mark_mode));
573 
574 #ifdef HXGE_DEBUG
575 		hpi_desc_handle.hxgep = hxgep;
576 		hpi_desc_handle.function.function = 0;
577 		hpi_desc_handle.function.instance = hxgep->instance;
578 		sad = save_desc_p->bits.sad;
579 		sad = (sad << 32) | save_desc_p->bits.sad_l;
580 		xfer_len = save_desc_p->bits.tr_len;
581 
582 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "\n\t: value 0x%llx\n"
583 		    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\t"
584 		    "mark %d sop %d\n",
585 		    save_desc_p->value, sad, save_desc_p->bits.tr_len,
586 		    xfer_len, save_desc_p->bits.num_ptr,
587 		    save_desc_p->bits.mark, save_desc_p->bits.sop));
588 
589 		hpi_txdma_dump_desc_one(hpi_desc_handle, NULL, i);
590 #endif
591 
592 		tx_msg_p->tx_msg_size = clen;
593 		i = TXDMA_DESC_NEXT_INDEX(i, 1, tx_ring_p->tx_wrap_mask);
594 		if (ngathers > hxge_tx_max_gathers) {
595 			good_packet = B_FALSE;
596 			mac_hcksum_get(mp, &start_offset, &stuff_offset,
597 			    &end_offset, &value, &cksum_flags);
598 
599 			HXGE_DEBUG_MSG((NULL, TX_CTL,
600 			    "==> hxge_start(14): pull msg - "
601 			    "len %d pkt_len %d ngathers %d",
602 			    len, pkt_len, ngathers));
603 			goto hxge_start_fail2;
604 		}
605 	} /* while (nmp) */
606 
607 	tx_msg_p->tx_message = mp;
608 	tx_desc_p = &tx_desc_ring_vp[sop_index];
609 #if defined(__i386)
610 	hpi_desc_handle.regp = (uint32_t)tx_desc_p;
611 #else
612 	hpi_desc_handle.regp = (uint64_t)tx_desc_p;
613 #endif
614 
615 	pkthdrp = (p_tx_pkt_hdr_all_t)hdrp;
616 	pkthdrp->reserved = 0;
617 	hdrp->value = 0;
618 	(void) hxge_fill_tx_hdr(mp, B_FALSE, cksum_on,
619 	    (pkt_len - TX_PKT_HEADER_SIZE), npads, pkthdrp);
620 
621 	/*
622 	 * Hardware header should not be counted as part of the frame
623 	 * when determining the frame size
624 	 */
625 	if ((pkt_len - TX_PKT_HEADER_SIZE) > (STD_FRAME_SIZE - ETHERFCSL)) {
626 		tdc_stats->tx_jumbo_pkts++;
627 	}
628 
629 	min_len = (hxgep->msg_min + TX_PKT_HEADER_SIZE + (npads * 2));
630 	if (pkt_len < min_len) {
631 		/* Assume we use bcopy to premapped buffers */
632 		kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
633 		HXGE_DEBUG_MSG((NULL, TX_CTL,
634 		    "==> hxge_start(14-1): < (msg_min + 16)"
635 		    "len %d pkt_len %d min_len %d bzero %d ngathers %d",
636 		    len, pkt_len, min_len, (min_len - pkt_len), ngathers));
637 		bzero((kaddr + pkt_len), (min_len - pkt_len));
638 		pkt_len = tx_msg_p->tx_msg_size = min_len;
639 
640 		sop_tx_desc_p->bits.tr_len = min_len;
641 
642 		HXGE_MEM_PIO_WRITE64(hpi_desc_handle, sop_tx_desc_p->value);
643 		tx_desc_p->value = sop_tx_desc_p->value;
644 
645 		HXGE_DEBUG_MSG((NULL, TX_CTL,
646 		    "==> hxge_start(14-2): < msg_min - "
647 		    "len %d pkt_len %d min_len %d ngathers %d",
648 		    len, pkt_len, min_len, ngathers));
649 	}
650 
651 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: cksum_flags 0x%x ",
652 	    cksum_flags));
653 	if (cksum_flags & HCK_PARTIALCKSUM) {
654 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
655 		    "==> hxge_start: cksum_flags 0x%x (partial checksum) ",
656 		    cksum_flags));
657 		cksum_on = B_TRUE;
658 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
659 		    "==> hxge_start: from IP cksum_flags 0x%x "
660 		    "(partial checksum) "
661 		    "start_offset %d stuff_offset %d",
662 		    cksum_flags, start_offset, stuff_offset));
663 		tmp_len = (uint64_t)(start_offset >> 1);
664 		hdrp->value |= (tmp_len << TX_PKT_HEADER_L4START_SHIFT);
665 		tmp_len = (uint64_t)(stuff_offset >> 1);
666 		hdrp->value |= (tmp_len << TX_PKT_HEADER_L4STUFF_SHIFT);
667 
668 		HXGE_DEBUG_MSG((hxgep, TX_CTL,
669 		    "==> hxge_start: from IP cksum_flags 0x%x "
670 		    "(partial checksum) "
671 		    "after SHIFT start_offset %d stuff_offset %d",
672 		    cksum_flags, start_offset, stuff_offset));
673 	}
674 
675 	/*
676 	 * pkt_len already includes 16 + paddings!!
677 	 * Update the control header length
678 	 */
679 
680 	/*
681 	 * Note that Hydra is different from Neptune where
682 	 * tot_xfer_len = (pkt_len - TX_PKT_HEADER_SIZE);
683 	 */
684 	tot_xfer_len = pkt_len;
685 	tmp_len = hdrp->value |
686 	    (tot_xfer_len << TX_PKT_HEADER_TOT_XFER_LEN_SHIFT);
687 
688 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
689 	    "==> hxge_start(15_x1): setting SOP "
690 	    "tot_xfer_len 0x%llx (%d) pkt_len %d tmp_len "
691 	    "0x%llx hdrp->value 0x%llx",
692 	    tot_xfer_len, tot_xfer_len, pkt_len, tmp_len, hdrp->value));
693 #if defined(_BIG_ENDIAN)
694 	hdrp->value = ddi_swap64(tmp_len);
695 #else
696 	hdrp->value = tmp_len;
697 #endif
698 	HXGE_DEBUG_MSG((hxgep,
699 	    TX_CTL, "==> hxge_start(15_x2): setting SOP "
700 	    "after SWAP: tot_xfer_len 0x%llx pkt_len %d "
701 	    "tmp_len 0x%llx hdrp->value 0x%llx",
702 	    tot_xfer_len, pkt_len, tmp_len, hdrp->value));
703 
704 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start(15): setting SOP "
705 	    "wr_index %d tot_xfer_len (%d) pkt_len %d npads %d",
706 	    sop_index, tot_xfer_len, pkt_len, npads));
707 
708 	sop_tx_desc_p->bits.sop = 1;
709 	sop_tx_desc_p->bits.mark = mark_mode;
710 	sop_tx_desc_p->bits.num_ptr = ngathers;
711 
712 	if (mark_mode)
713 		tdc_stats->tx_marks++;
714 
715 	HXGE_MEM_PIO_WRITE64(hpi_desc_handle, sop_tx_desc_p->value);
716 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start(16): set SOP done"));
717 
718 #ifdef HXGE_DEBUG
719 	hpi_desc_handle.hxgep = hxgep;
720 	hpi_desc_handle.function.function = 0;
721 	hpi_desc_handle.function.instance = hxgep->instance;
722 
723 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "\n\t: value 0x%llx\n"
724 	    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\tmark %d sop %d\n",
725 	    save_desc_p->value, sad, save_desc_p->bits.tr_len,
726 	    xfer_len, save_desc_p->bits.num_ptr, save_desc_p->bits.mark,
727 	    save_desc_p->bits.sop));
728 	(void) hpi_txdma_dump_desc_one(hpi_desc_handle, NULL, sop_index);
729 
730 	dump_len = (pkt_len > 128) ? 128: pkt_len;
731 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
732 	    "==> hxge_start: dump packets(17) (after sop set, len "
733 	    " (len/dump_len/pkt_len/tot_xfer_len) %d/%d/%d/%d):\n"
734 	    "ptr $%p: %s", len, dump_len, pkt_len, tot_xfer_len,
735 	    (char *)hdrp, hxge_dump_packet((char *)hdrp, dump_len)));
736 	HXGE_DEBUG_MSG((hxgep, TX_CTL,
737 	    "==> hxge_start(18): TX desc sync: sop_index %d", sop_index));
738 #endif
739 
740 	if ((ngathers == 1) || tx_ring_p->wr_index < i) {
741 		(void) ddi_dma_sync(tx_desc_dma_handle,
742 		    sop_index * sizeof (tx_desc_t),
743 		    ngathers * sizeof (tx_desc_t), DDI_DMA_SYNC_FORDEV);
744 
745 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "hxge_start(19): sync 1 "
746 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
747 		    "pkt_len %d ngathers %d sop_index %d\n",
748 		    stuff_offset, start_offset,
749 		    pkt_len, ngathers, sop_index));
750 	} else { /* more than one descriptor and wrap around */
751 		uint32_t nsdescs = tx_ring_p->tx_ring_size - sop_index;
752 		(void) ddi_dma_sync(tx_desc_dma_handle,
753 		    sop_index * sizeof (tx_desc_t),
754 		    nsdescs * sizeof (tx_desc_t), DDI_DMA_SYNC_FORDEV);
755 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "hxge_start(20): sync 1 "
756 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
757 		    "pkt_len %d ngathers %d sop_index %d\n",
758 		    stuff_offset, start_offset, pkt_len, ngathers, sop_index));
759 
760 		(void) ddi_dma_sync(tx_desc_dma_handle, 0,
761 		    (ngathers - nsdescs) * sizeof (tx_desc_t),
762 		    DDI_DMA_SYNC_FORDEV);
763 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "hxge_start(21): sync 2 "
764 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
765 		    "pkt_len %d ngathers %d sop_index %d\n",
766 		    stuff_offset, start_offset,
767 		    pkt_len, ngathers, sop_index));
768 	}
769 
770 	tail_index = tx_ring_p->wr_index;
771 	tail_wrap = tx_ring_p->wr_index_wrap;
772 
773 	tx_ring_p->wr_index = i;
774 	if (tx_ring_p->wr_index <= tail_index) {
775 		tx_ring_p->wr_index_wrap = ((tail_wrap == B_TRUE) ?
776 		    B_FALSE : B_TRUE);
777 	}
778 
779 	tx_ring_p->descs_pending += ngathers;
780 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: TX kick: "
781 	    "channel %d wr_index %d wrap %d ngathers %d desc_pend %d",
782 	    tx_ring_p->tdc, tx_ring_p->wr_index, tx_ring_p->wr_index_wrap,
783 	    ngathers, tx_ring_p->descs_pending));
784 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: TX KICKING: "));
785 
786 	kick.value = 0;
787 	kick.bits.wrap = tx_ring_p->wr_index_wrap;
788 	kick.bits.tail = (uint16_t)tx_ring_p->wr_index;
789 
790 	/* Kick start the Transmit kick register */
791 	TXDMA_REG_WRITE64(HXGE_DEV_HPI_HANDLE(hxgep),
792 	    TDC_TDR_KICK, (uint8_t)tx_ring_p->tdc, kick.value);
793 	tdc_stats->tx_starts++;
794 	MUTEX_EXIT(&tx_ring_p->lock);
795 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "<== hxge_start"));
796 	return (status);
797 
798 hxge_start_fail2:
799 	if (good_packet == B_FALSE) {
800 		cur_index = sop_index;
801 		HXGE_DEBUG_MSG((hxgep, TX_CTL, "==> hxge_start: clean up"));
802 		for (i = 0; i < ngathers; i++) {
803 			tx_desc_p = &tx_desc_ring_vp[cur_index];
804 #if defined(__i386)
805 			hpi_handle.regp = (uint32_t)tx_desc_p;
806 #else
807 			hpi_handle.regp = (uint64_t)tx_desc_p;
808 #endif
809 			tx_msg_p = &tx_msg_ring[cur_index];
810 			(void) hpi_txdma_desc_set_zero(hpi_handle, 1);
811 			if (tx_msg_p->flags.dma_type == USE_DVMA) {
812 				HXGE_DEBUG_MSG((hxgep, TX_CTL,
813 				    "tx_desc_p = %X index = %d",
814 				    tx_desc_p, tx_ring_p->rd_index));
815 				(void) dvma_unload(tx_msg_p->dvma_handle,
816 				    0, -1);
817 				tx_msg_p->dvma_handle = NULL;
818 				if (tx_ring_p->dvma_wr_index ==
819 				    tx_ring_p->dvma_wrap_mask)
820 					tx_ring_p->dvma_wr_index = 0;
821 				else
822 					tx_ring_p->dvma_wr_index++;
823 				tx_ring_p->dvma_pending--;
824 			} else if (tx_msg_p->flags.dma_type == USE_DMA) {
825 				if (ddi_dma_unbind_handle(
826 				    tx_msg_p->dma_handle)) {
827 					cmn_err(CE_WARN, "hxge_start: "
828 					    "ddi_dma_unbind_handle failed");
829 				}
830 			}
831 			tx_msg_p->flags.dma_type = USE_NONE;
832 			cur_index = TXDMA_DESC_NEXT_INDEX(cur_index, 1,
833 			    tx_ring_p->tx_wrap_mask);
834 
835 		}
836 	}
837 
838 	MUTEX_EXIT(&tx_ring_p->lock);
839 
840 hxge_start_fail1:
841 	/* Add FMA to check the access handle hxge_hregh */
842 	HXGE_DEBUG_MSG((hxgep, TX_CTL, "<== hxge_start"));
843 	return (status);
844 }
845