xref: /illumos-gate/usr/src/uts/common/io/e1000g/e1000g_rx.c (revision da6c28aaf62fa55f0fdb8004aa40f88f23bf53f0)
1 /*
2  * This file is provided under a CDDLv1 license.  When using or
3  * redistributing this file, you may do so under this license.
4  * In redistributing this file this license must be included
5  * and no other modification of this header file is permitted.
6  *
7  * CDDL LICENSE SUMMARY
8  *
9  * Copyright(c) 1999 - 2007 Intel Corporation. All rights reserved.
10  *
11  * The contents of this file are subject to the terms of Version
12  * 1.0 of the Common Development and Distribution License (the "License").
13  *
14  * You should have received a copy of the License with this software.
15  * You can obtain a copy of the License at
16  *	http://www.opensolaris.org/os/licensing.
17  * See the License for the specific language governing permissions
18  * and limitations under the License.
19  */
20 
21 /*
22  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms of the CDDLv1.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * **********************************************************************
30  *									*
31  * Module Name:								*
32  *   e1000g_rx.c							*
33  *									*
34  * Abstract:								*
35  *   This file contains some routines that take care of Receive		*
36  *   interrupt and also for the received packets it sends up to		*
37  *   upper layer.							*
38  *   It tries to do a zero copy if free buffers are available in	*
39  *   the pool.								*
40  *									*
41  * **********************************************************************
42  */
43 
44 #include "e1000g_sw.h"
45 #include "e1000g_debug.h"
46 
47 static p_rx_sw_packet_t e1000g_get_buf(e1000g_rx_ring_t *rx_ring);
48 #pragma	inline(e1000g_get_buf)
49 static void e1000g_priv_devi_list_clean();
50 
51 /*
52  * e1000g_rxfree_func - the call-back function to reclaim rx buffer
53  *
54  * This function is called when an mp is freed by the user thru
55  * freeb call (Only for mp constructed through desballoc call)
56  * It returns back the freed buffer to the freelist
57  */
58 void
59 e1000g_rxfree_func(p_rx_sw_packet_t packet)
60 {
61 	struct e1000g *Adapter;
62 	e1000g_rx_ring_t *rx_ring;
63 
64 	rx_ring = (e1000g_rx_ring_t *)packet->rx_ring;
65 	Adapter = rx_ring->adapter;
66 
67 	/*
68 	 * Here the rx recycling processes different rx packets in different
69 	 * threads, so we protect it with RW_READER to ensure it won't block
70 	 * other rx recycling threads.
71 	 */
72 	rw_enter(&e1000g_rx_detach_lock, RW_READER);
73 
74 	if (packet->flag == E1000G_RX_SW_FREE) {
75 		rw_exit(&e1000g_rx_detach_lock);
76 		return;
77 	}
78 
79 	if (packet->flag == E1000G_RX_SW_STOP) {
80 		packet->flag = E1000G_RX_SW_FREE;
81 		rw_exit(&e1000g_rx_detach_lock);
82 
83 		rw_enter(&e1000g_rx_detach_lock, RW_WRITER);
84 		rx_ring->pending_count--;
85 		e1000g_mblks_pending--;
86 
87 		if (rx_ring->pending_count == 0) {
88 			while (rx_ring->pending_list != NULL) {
89 				packet = rx_ring->pending_list;
90 				rx_ring->pending_list =
91 				    rx_ring->pending_list->next;
92 
93 				ASSERT(packet->mp == NULL);
94 				e1000g_free_rx_sw_packet(packet);
95 			}
96 		}
97 
98 		/*
99 		 * If e1000g_force_detach is enabled, we need to clean up
100 		 * the idle priv_dip entries in the private dip list while
101 		 * e1000g_mblks_pending is zero.
102 		 */
103 		if (e1000g_force_detach && (e1000g_mblks_pending == 0))
104 			e1000g_priv_devi_list_clean();
105 		rw_exit(&e1000g_rx_detach_lock);
106 		return;
107 	}
108 
109 	if (packet->flag == E1000G_RX_SW_DETACH) {
110 		packet->flag = E1000G_RX_SW_FREE;
111 		rw_exit(&e1000g_rx_detach_lock);
112 
113 		ASSERT(packet->mp == NULL);
114 		e1000g_free_rx_sw_packet(packet);
115 
116 		/*
117 		 * Here the e1000g_mblks_pending may be modified by different
118 		 * rx recycling threads simultaneously, so we need to protect
119 		 * it with RW_WRITER.
120 		 */
121 		rw_enter(&e1000g_rx_detach_lock, RW_WRITER);
122 		e1000g_mblks_pending--;
123 
124 		/*
125 		 * If e1000g_force_detach is enabled, we need to clean up
126 		 * the idle priv_dip entries in the private dip list while
127 		 * e1000g_mblks_pending is zero.
128 		 */
129 		if (e1000g_force_detach && (e1000g_mblks_pending == 0))
130 			e1000g_priv_devi_list_clean();
131 		rw_exit(&e1000g_rx_detach_lock);
132 		return;
133 	}
134 
135 	packet->flag = E1000G_RX_SW_FREE;
136 
137 	if (packet->mp == NULL) {
138 		/*
139 		 * Allocate a mblk that binds to the data buffer
140 		 */
141 		packet->mp = desballoc((unsigned char *)
142 		    packet->rx_buf->address - E1000G_IPALIGNROOM,
143 		    packet->rx_buf->size + E1000G_IPALIGNROOM,
144 		    BPRI_MED, &packet->free_rtn);
145 
146 		if (packet->mp != NULL) {
147 			packet->mp->b_rptr += E1000G_IPALIGNROOM;
148 			packet->mp->b_wptr += E1000G_IPALIGNROOM;
149 		} else {
150 			E1000G_STAT(rx_ring->stat_esballoc_fail);
151 		}
152 	}
153 
154 	mutex_enter(&rx_ring->freelist_lock);
155 	QUEUE_PUSH_TAIL(&rx_ring->free_list, &packet->Link);
156 	rx_ring->avail_freepkt++;
157 	mutex_exit(&rx_ring->freelist_lock);
158 
159 	rw_exit(&e1000g_rx_detach_lock);
160 }
161 
162 /*
163  * e1000g_priv_devi_list_clean - clean up e1000g_private_devi_list
164  *
165  * We will walk the e1000g_private_devi_list to free the entry marked
166  * with the E1000G_PRIV_DEVI_DETACH flag.
167  */
168 static void
169 e1000g_priv_devi_list_clean()
170 {
171 	private_devi_list_t *devi_node, *devi_del;
172 
173 	if (e1000g_private_devi_list == NULL)
174 		return;
175 
176 	devi_node = e1000g_private_devi_list;
177 	while ((devi_node != NULL) &&
178 	    (devi_node->flag == E1000G_PRIV_DEVI_DETACH)) {
179 		e1000g_private_devi_list = devi_node->next;
180 		kmem_free(devi_node->priv_dip,
181 		    sizeof (struct dev_info));
182 		kmem_free(devi_node,
183 		    sizeof (private_devi_list_t));
184 		devi_node = e1000g_private_devi_list;
185 	}
186 	if (e1000g_private_devi_list == NULL)
187 		return;
188 	while (devi_node->next != NULL) {
189 		if (devi_node->next->flag == E1000G_PRIV_DEVI_DETACH) {
190 			devi_del = devi_node->next;
191 			devi_node->next = devi_del->next;
192 			kmem_free(devi_del->priv_dip,
193 			    sizeof (struct dev_info));
194 			kmem_free(devi_del,
195 			    sizeof (private_devi_list_t));
196 		} else {
197 			devi_node = devi_node->next;
198 		}
199 	}
200 }
201 
202 /*
203  * e1000g_rx_setup - setup rx data structures
204  *
205  * This routine initializes all of the receive related
206  * structures. This includes the receive descriptors, the
207  * actual receive buffers, and the rx_sw_packet software
208  * structures.
209  */
210 void
211 e1000g_rx_setup(struct e1000g *Adapter)
212 {
213 	struct e1000_hw *hw;
214 	p_rx_sw_packet_t packet;
215 	struct e1000_rx_desc *descriptor;
216 	uint32_t buf_low;
217 	uint32_t buf_high;
218 	uint32_t reg_val;
219 	int i;
220 	int size;
221 	e1000g_rx_ring_t *rx_ring;
222 
223 	hw = &Adapter->shared;
224 	rx_ring = Adapter->rx_ring;
225 
226 	/*
227 	 * zero out all of the receive buffer descriptor memory
228 	 * assures any previous data or status is erased
229 	 */
230 	bzero(rx_ring->rbd_area,
231 	    sizeof (struct e1000_rx_desc) * Adapter->rx_desc_num);
232 
233 	if (!Adapter->rx_buffer_setup) {
234 		/* Init the list of "Receive Buffer" */
235 		QUEUE_INIT_LIST(&rx_ring->recv_list);
236 
237 		/* Init the list of "Free Receive Buffer" */
238 		QUEUE_INIT_LIST(&rx_ring->free_list);
239 
240 		/*
241 		 * Setup Receive list and the Free list. Note that
242 		 * the both were allocated in one packet area.
243 		 */
244 		packet = rx_ring->packet_area;
245 		descriptor = rx_ring->rbd_first;
246 
247 		for (i = 0; i < Adapter->rx_desc_num;
248 		    i++, packet = packet->next, descriptor++) {
249 			ASSERT(packet != NULL);
250 			ASSERT(descriptor != NULL);
251 			descriptor->buffer_addr =
252 			    packet->rx_buf->dma_address;
253 
254 			/* Add this rx_sw_packet to the receive list */
255 			QUEUE_PUSH_TAIL(&rx_ring->recv_list,
256 			    &packet->Link);
257 		}
258 
259 		for (i = 0; i < Adapter->rx_freelist_num;
260 		    i++, packet = packet->next) {
261 			ASSERT(packet != NULL);
262 			/* Add this rx_sw_packet to the free list */
263 			QUEUE_PUSH_TAIL(&rx_ring->free_list,
264 			    &packet->Link);
265 		}
266 		rx_ring->avail_freepkt = Adapter->rx_freelist_num;
267 
268 		Adapter->rx_buffer_setup = B_TRUE;
269 	} else {
270 		/* Setup the initial pointer to the first rx descriptor */
271 		packet = (p_rx_sw_packet_t)
272 		    QUEUE_GET_HEAD(&rx_ring->recv_list);
273 		descriptor = rx_ring->rbd_first;
274 
275 		for (i = 0; i < Adapter->rx_desc_num; i++) {
276 			ASSERT(packet != NULL);
277 			ASSERT(descriptor != NULL);
278 			descriptor->buffer_addr =
279 			    packet->rx_buf->dma_address;
280 
281 			/* Get next rx_sw_packet */
282 			packet = (p_rx_sw_packet_t)
283 			    QUEUE_GET_NEXT(&rx_ring->recv_list, &packet->Link);
284 			descriptor++;
285 		}
286 	}
287 
288 	/*
289 	 * Setup our descriptor pointers
290 	 */
291 	rx_ring->rbd_next = rx_ring->rbd_first;
292 
293 	size = Adapter->rx_desc_num * sizeof (struct e1000_rx_desc);
294 	E1000_WRITE_REG(hw, E1000_RDLEN, size);
295 	size = E1000_READ_REG(hw, E1000_RDLEN);
296 
297 	/* To get lower order bits */
298 	buf_low = (uint32_t)rx_ring->rbd_dma_addr;
299 	/* To get the higher order bits */
300 	buf_high = (uint32_t)(rx_ring->rbd_dma_addr >> 32);
301 
302 	E1000_WRITE_REG(hw, E1000_RDBAH, buf_high);
303 	E1000_WRITE_REG(hw, E1000_RDBAL, buf_low);
304 
305 	/*
306 	 * Setup our HW Rx Head & Tail descriptor pointers
307 	 */
308 	E1000_WRITE_REG(hw, E1000_RDT,
309 	    (uint32_t)(rx_ring->rbd_last - rx_ring->rbd_first));
310 	E1000_WRITE_REG(hw, E1000_RDH, 0);
311 
312 	/*
313 	 * Setup the Receive Control Register (RCTL), and ENABLE the
314 	 * receiver. The initial configuration is to: Enable the receiver,
315 	 * accept broadcasts, discard bad packets (and long packets),
316 	 * disable VLAN filter checking, set the receive descriptor
317 	 * minimum threshold size to 1/2, and the receive buffer size to
318 	 * 2k.
319 	 */
320 	reg_val = E1000_RCTL_EN |	/* Enable Receive Unit */
321 	    E1000_RCTL_BAM |		/* Accept Broadcast Packets */
322 	    E1000_RCTL_LPE |		/* Large Packet Enable bit */
323 	    (hw->mac.mc_filter_type << E1000_RCTL_MO_SHIFT) |
324 	    E1000_RCTL_RDMTS_HALF |
325 	    E1000_RCTL_LBM_NO;		/* Loopback Mode = none */
326 
327 	if (Adapter->strip_crc)
328 		reg_val |= E1000_RCTL_SECRC;	/* Strip Ethernet CRC */
329 
330 	switch (hw->mac.max_frame_size) {
331 	case ETHERMAX:
332 		reg_val |= E1000_RCTL_SZ_2048;
333 		break;
334 	case FRAME_SIZE_UPTO_4K:
335 		reg_val |= E1000_RCTL_SZ_4096 | E1000_RCTL_BSEX;
336 		break;
337 	case FRAME_SIZE_UPTO_8K:
338 		reg_val |= E1000_RCTL_SZ_8192 | E1000_RCTL_BSEX;
339 		break;
340 	case FRAME_SIZE_UPTO_9K:
341 	case FRAME_SIZE_UPTO_16K:
342 		reg_val |= E1000_RCTL_SZ_16384 | E1000_RCTL_BSEX;
343 		break;
344 	default:
345 		reg_val |= E1000_RCTL_SZ_2048;
346 		break;
347 	}
348 
349 	if (e1000_tbi_sbp_enabled_82543(hw))
350 		reg_val |= E1000_RCTL_SBP;
351 
352 	/*
353 	 * Enable early receives on supported devices, only takes effect when
354 	 * packet size is equal or larger than the specified value (in 8 byte
355 	 * units), e.g. using jumbo frames when setting to E1000_ERT_2048
356 	 */
357 	if ((hw->mac.type == e1000_82573) || (hw->mac.type == e1000_ich9lan))
358 		E1000_WRITE_REG(hw, E1000_ERT, E1000_ERT_2048);
359 
360 	E1000_WRITE_REG(hw, E1000_RCTL, reg_val);
361 
362 	reg_val =
363 	    E1000_RXCSUM_TUOFL |	/* TCP/UDP checksum offload Enable */
364 	    E1000_RXCSUM_IPOFL;		/* IP checksum offload Enable */
365 
366 	E1000_WRITE_REG(hw, E1000_RXCSUM, reg_val);
367 }
368 
369 /*
370  * e1000g_get_buf - get an rx sw packet from the free_list
371  */
372 static p_rx_sw_packet_t
373 e1000g_get_buf(e1000g_rx_ring_t *rx_ring)
374 {
375 	struct e1000g *Adapter;
376 	p_rx_sw_packet_t packet;
377 
378 	Adapter = rx_ring->adapter;
379 
380 	mutex_enter(&rx_ring->freelist_lock);
381 	packet = (p_rx_sw_packet_t)
382 	    QUEUE_POP_HEAD(&rx_ring->free_list);
383 	if (packet != NULL)
384 		rx_ring->avail_freepkt--;
385 	mutex_exit(&rx_ring->freelist_lock);
386 
387 	return (packet);
388 }
389 
390 /*
391  * e1000g_receive - main receive routine
392  *
393  * This routine will process packets received in an interrupt
394  */
395 mblk_t *
396 e1000g_receive(struct e1000g *Adapter)
397 {
398 	struct e1000_hw *hw;
399 	mblk_t *nmp;
400 	mblk_t *ret_mp;
401 	mblk_t *ret_nmp;
402 	struct e1000_rx_desc *current_desc;
403 	struct e1000_rx_desc *last_desc;
404 	p_rx_sw_packet_t packet;
405 	p_rx_sw_packet_t newpkt;
406 	USHORT length;
407 	uint32_t pkt_count;
408 	uint32_t desc_count;
409 	boolean_t accept_frame;
410 	boolean_t end_of_packet;
411 	boolean_t need_copy;
412 	e1000g_rx_ring_t *rx_ring;
413 	dma_buffer_t *rx_buf;
414 	uint16_t cksumflags;
415 
416 	ret_mp = NULL;
417 	ret_nmp = NULL;
418 	pkt_count = 0;
419 	desc_count = 0;
420 	cksumflags = 0;
421 
422 	hw = &Adapter->shared;
423 	rx_ring = Adapter->rx_ring;
424 
425 	/* Sync the Rx descriptor DMA buffers */
426 	(void) ddi_dma_sync(rx_ring->rbd_dma_handle,
427 	    0, 0, DDI_DMA_SYNC_FORKERNEL);
428 
429 	if (e1000g_check_dma_handle(rx_ring->rbd_dma_handle) != DDI_FM_OK) {
430 		ddi_fm_service_impact(Adapter->dip, DDI_SERVICE_DEGRADED);
431 		Adapter->chip_state = E1000G_ERROR;
432 	}
433 
434 	current_desc = rx_ring->rbd_next;
435 	if (!(current_desc->status & E1000_RXD_STAT_DD)) {
436 		/*
437 		 * don't send anything up. just clear the RFD
438 		 */
439 		E1000G_DEBUG_STAT(rx_ring->stat_none);
440 		return (ret_mp);
441 	}
442 
443 	/*
444 	 * Loop through the receive descriptors starting at the last known
445 	 * descriptor owned by the hardware that begins a packet.
446 	 */
447 	while ((current_desc->status & E1000_RXD_STAT_DD) &&
448 	    (pkt_count < Adapter->rx_limit_onintr)) {
449 
450 		desc_count++;
451 		/*
452 		 * Now this can happen in Jumbo frame situation.
453 		 */
454 		if (current_desc->status & E1000_RXD_STAT_EOP) {
455 			/* packet has EOP set */
456 			end_of_packet = B_TRUE;
457 		} else {
458 			/*
459 			 * If this received buffer does not have the
460 			 * End-Of-Packet bit set, the received packet
461 			 * will consume multiple buffers. We won't send this
462 			 * packet upstack till we get all the related buffers.
463 			 */
464 			end_of_packet = B_FALSE;
465 		}
466 
467 		/*
468 		 * Get a pointer to the actual receive buffer
469 		 * The mp->b_rptr is mapped to The CurrentDescriptor
470 		 * Buffer Address.
471 		 */
472 		packet =
473 		    (p_rx_sw_packet_t)QUEUE_GET_HEAD(&rx_ring->recv_list);
474 		ASSERT(packet != NULL);
475 
476 		rx_buf = packet->rx_buf;
477 
478 		length = current_desc->length;
479 
480 #ifdef __sparc
481 		if (packet->dma_type == USE_DVMA)
482 			dvma_sync(rx_buf->dma_handle, 0,
483 			    DDI_DMA_SYNC_FORKERNEL);
484 		else
485 			(void) ddi_dma_sync(rx_buf->dma_handle,
486 			    E1000G_IPALIGNROOM, length,
487 			    DDI_DMA_SYNC_FORKERNEL);
488 #else
489 		(void) ddi_dma_sync(rx_buf->dma_handle,
490 		    E1000G_IPALIGNROOM, length,
491 		    DDI_DMA_SYNC_FORKERNEL);
492 #endif
493 
494 		if (e1000g_check_dma_handle(
495 		    rx_buf->dma_handle) != DDI_FM_OK) {
496 			ddi_fm_service_impact(Adapter->dip,
497 			    DDI_SERVICE_DEGRADED);
498 			Adapter->chip_state = E1000G_ERROR;
499 		}
500 
501 		accept_frame = (current_desc->errors == 0) ||
502 		    ((current_desc->errors &
503 		    (E1000_RXD_ERR_TCPE | E1000_RXD_ERR_IPE)) != 0);
504 
505 		if (hw->mac.type == e1000_82543) {
506 			unsigned char last_byte;
507 
508 			last_byte =
509 			    *((unsigned char *)rx_buf->address + length - 1);
510 
511 			if (TBI_ACCEPT(hw,
512 			    current_desc->status, current_desc->errors,
513 			    current_desc->length, last_byte)) {
514 
515 				e1000_tbi_adjust_stats(Adapter,
516 				    length, hw->mac.addr);
517 
518 				length--;
519 				accept_frame = B_TRUE;
520 			} else if (e1000_tbi_sbp_enabled_82543(hw) &&
521 			    (current_desc->errors == E1000_RXD_ERR_CE)) {
522 				accept_frame = B_TRUE;
523 			}
524 		}
525 
526 		/*
527 		 * Indicate the packet to the NOS if it was good.
528 		 * Normally, hardware will discard bad packets for us.
529 		 * Check for the packet to be a valid Ethernet packet
530 		 */
531 		if (!accept_frame) {
532 			/*
533 			 * error in incoming packet, either the packet is not a
534 			 * ethernet size packet, or the packet has an error. In
535 			 * either case, the packet will simply be discarded.
536 			 */
537 			E1000G_DEBUGLOG_0(Adapter, E1000G_INFO_LEVEL,
538 			    "Process Receive Interrupts: Error in Packet\n");
539 
540 			E1000G_STAT(rx_ring->stat_error);
541 			/*
542 			 * Returning here as we are done here. There is
543 			 * no point in waiting for while loop to elapse
544 			 * and the things which were done. More efficient
545 			 * and less error prone...
546 			 */
547 			goto rx_drop;
548 		}
549 
550 		/*
551 		 * If the Ethernet CRC is not stripped by the hardware,
552 		 * we need to strip it before sending it up to the stack.
553 		 */
554 		if (end_of_packet && !Adapter->strip_crc) {
555 			if (length > CRC_LENGTH) {
556 				length -= CRC_LENGTH;
557 			} else {
558 				/*
559 				 * If the fragment is smaller than the CRC,
560 				 * drop this fragment, do the processing of
561 				 * the end of the packet.
562 				 */
563 				ASSERT(rx_ring->rx_mblk_tail != NULL);
564 				rx_ring->rx_mblk_tail->b_wptr -=
565 				    CRC_LENGTH - length;
566 				rx_ring->rx_mblk_len -=
567 				    CRC_LENGTH - length;
568 
569 				QUEUE_POP_HEAD(&rx_ring->recv_list);
570 
571 				goto rx_end_of_packet;
572 			}
573 		}
574 
575 		need_copy = B_TRUE;
576 
577 		if (length <= Adapter->rx_bcopy_thresh)
578 			goto rx_copy;
579 
580 		/*
581 		 * Get the pre-constructed mblk that was associated
582 		 * to the receive data buffer.
583 		 */
584 		if (packet->mp == NULL) {
585 			packet->mp = desballoc((unsigned char *)
586 			    rx_buf->address - E1000G_IPALIGNROOM,
587 			    length + E1000G_IPALIGNROOM,
588 			    BPRI_MED, &packet->free_rtn);
589 
590 			if (packet->mp != NULL) {
591 				packet->mp->b_rptr += E1000G_IPALIGNROOM;
592 				packet->mp->b_wptr += E1000G_IPALIGNROOM;
593 			} else {
594 				E1000G_STAT(rx_ring->stat_esballoc_fail);
595 			}
596 		}
597 
598 		if (packet->mp != NULL) {
599 			/*
600 			 * We have two sets of buffer pool. One associated with
601 			 * the Rxdescriptors and other a freelist buffer pool.
602 			 * Each time we get a good packet, Try to get a buffer
603 			 * from the freelist pool using e1000g_get_buf. If we
604 			 * get free buffer, then replace the descriptor buffer
605 			 * address with the free buffer we just got, and pass
606 			 * the pre-constructed mblk upstack. (note no copying)
607 			 *
608 			 * If we failed to get a free buffer, then try to
609 			 * allocate a new buffer(mp) and copy the recv buffer
610 			 * content to our newly allocated buffer(mp). Don't
611 			 * disturb the desriptor buffer address. (note copying)
612 			 */
613 			newpkt = e1000g_get_buf(rx_ring);
614 
615 			if (newpkt != NULL) {
616 				/*
617 				 * Get the mblk associated to the data,
618 				 * and strip it off the sw packet.
619 				 */
620 				nmp = packet->mp;
621 				packet->mp = NULL;
622 				packet->flag = E1000G_RX_SW_SENDUP;
623 
624 				/*
625 				 * Now replace old buffer with the new
626 				 * one we got from free list
627 				 * Both the RxSwPacket as well as the
628 				 * Receive Buffer Descriptor will now
629 				 * point to this new packet.
630 				 */
631 				packet = newpkt;
632 
633 				current_desc->buffer_addr =
634 				    newpkt->rx_buf->dma_address;
635 
636 				need_copy = B_FALSE;
637 			} else {
638 				E1000G_DEBUG_STAT(rx_ring->stat_no_freepkt);
639 			}
640 		}
641 
642 rx_copy:
643 		if (need_copy) {
644 			/*
645 			 * No buffers available on free list,
646 			 * bcopy the data from the buffer and
647 			 * keep the original buffer. Dont want to
648 			 * do this.. Yack but no other way
649 			 */
650 			if ((nmp = allocb(length + E1000G_IPALIGNROOM,
651 			    BPRI_MED)) == NULL) {
652 				/*
653 				 * The system has no buffers available
654 				 * to send up the incoming packet, hence
655 				 * the packet will have to be processed
656 				 * when there're more buffers available.
657 				 */
658 				E1000G_STAT(rx_ring->stat_allocb_fail);
659 				goto rx_drop;
660 			}
661 			nmp->b_rptr += E1000G_IPALIGNROOM;
662 			nmp->b_wptr += E1000G_IPALIGNROOM;
663 			/*
664 			 * The free list did not have any buffers
665 			 * available, so, the received packet will
666 			 * have to be copied into a mp and the original
667 			 * buffer will have to be retained for future
668 			 * packet reception.
669 			 */
670 			bcopy(rx_buf->address, nmp->b_wptr, length);
671 		}
672 
673 		/*
674 		 * The rx_sw_packet MUST be popped off the
675 		 * RxSwPacketList before either a putnext or freemsg
676 		 * is done on the mp that has now been created by the
677 		 * desballoc. If not, it is possible that the free
678 		 * routine will get called from the interrupt context
679 		 * and try to put this packet on the free list
680 		 */
681 		(p_rx_sw_packet_t)QUEUE_POP_HEAD(&rx_ring->recv_list);
682 
683 		ASSERT(nmp != NULL);
684 		nmp->b_wptr += length;
685 
686 		if (rx_ring->rx_mblk == NULL) {
687 			/*
688 			 *  TCP/UDP checksum offload and
689 			 *  IP checksum offload
690 			 */
691 			if (!(current_desc->status & E1000_RXD_STAT_IXSM)) {
692 				/*
693 				 * Check TCP/UDP checksum
694 				 */
695 				if ((current_desc->status &
696 				    E1000_RXD_STAT_TCPCS) &&
697 				    !(current_desc->errors &
698 				    E1000_RXD_ERR_TCPE))
699 					cksumflags |= HCK_FULLCKSUM |
700 					    HCK_FULLCKSUM_OK;
701 				/*
702 				 * Check IP Checksum
703 				 */
704 				if ((current_desc->status &
705 				    E1000_RXD_STAT_IPCS) &&
706 				    !(current_desc->errors &
707 				    E1000_RXD_ERR_IPE))
708 					cksumflags |= HCK_IPV4_HDRCKSUM;
709 			}
710 		}
711 
712 		/*
713 		 * We need to maintain our packet chain in the global
714 		 * Adapter structure, for the Rx processing can end
715 		 * with a fragment that has no EOP set.
716 		 */
717 		if (rx_ring->rx_mblk == NULL) {
718 			/* Get the head of the message chain */
719 			rx_ring->rx_mblk = nmp;
720 			rx_ring->rx_mblk_tail = nmp;
721 			rx_ring->rx_mblk_len = length;
722 		} else {	/* Not the first packet */
723 			/* Continue adding buffers */
724 			rx_ring->rx_mblk_tail->b_cont = nmp;
725 			rx_ring->rx_mblk_tail = nmp;
726 			rx_ring->rx_mblk_len += length;
727 		}
728 		ASSERT(rx_ring->rx_mblk != NULL);
729 		ASSERT(rx_ring->rx_mblk_tail != NULL);
730 		ASSERT(rx_ring->rx_mblk_tail->b_cont == NULL);
731 
732 		/*
733 		 * Now this MP is ready to travel upwards but some more
734 		 * fragments are coming.
735 		 * We will send packet upwards as soon as we get EOP
736 		 * set on the packet.
737 		 */
738 		if (!end_of_packet) {
739 			/*
740 			 * continue to get the next descriptor,
741 			 * Tail would be advanced at the end
742 			 */
743 			goto rx_next_desc;
744 		}
745 
746 rx_end_of_packet:
747 		/*
748 		 * Found packet with EOP
749 		 * Process the last fragment.
750 		 */
751 		if (cksumflags != 0) {
752 			(void) hcksum_assoc(rx_ring->rx_mblk,
753 			    NULL, NULL, 0, 0, 0, 0, cksumflags, 0);
754 			cksumflags = 0;
755 		}
756 
757 		/*
758 		 * Count packets that span multi-descriptors
759 		 */
760 		E1000G_DEBUG_STAT_COND(rx_ring->stat_multi_desc,
761 		    (rx_ring->rx_mblk->b_cont != NULL));
762 
763 		/*
764 		 * Append to list to send upstream
765 		 */
766 		if (ret_mp == NULL) {
767 			ret_mp = ret_nmp = rx_ring->rx_mblk;
768 		} else {
769 			ret_nmp->b_next = rx_ring->rx_mblk;
770 			ret_nmp = rx_ring->rx_mblk;
771 		}
772 		ret_nmp->b_next = NULL;
773 
774 		rx_ring->rx_mblk = NULL;
775 		rx_ring->rx_mblk_tail = NULL;
776 		rx_ring->rx_mblk_len = 0;
777 
778 		pkt_count++;
779 
780 rx_next_desc:
781 		/*
782 		 * Zero out the receive descriptors status
783 		 */
784 		current_desc->status = 0;
785 
786 		if (current_desc == rx_ring->rbd_last)
787 			rx_ring->rbd_next = rx_ring->rbd_first;
788 		else
789 			rx_ring->rbd_next++;
790 
791 		last_desc = current_desc;
792 		current_desc = rx_ring->rbd_next;
793 
794 		/*
795 		 * Put the buffer that we just indicated back
796 		 * at the end of our list
797 		 */
798 		QUEUE_PUSH_TAIL(&rx_ring->recv_list,
799 		    &packet->Link);
800 	}	/* while loop */
801 
802 	if (pkt_count >= Adapter->rx_limit_onintr)
803 		E1000G_STAT(rx_ring->stat_exceed_pkt);
804 
805 	/* Sync the Rx descriptor DMA buffers */
806 	(void) ddi_dma_sync(rx_ring->rbd_dma_handle,
807 	    0, 0, DDI_DMA_SYNC_FORDEV);
808 
809 	/*
810 	 * Advance the E1000's Receive Queue #0 "Tail Pointer".
811 	 */
812 	E1000_WRITE_REG(hw, E1000_RDT,
813 	    (uint32_t)(last_desc - rx_ring->rbd_first));
814 
815 	if (e1000g_check_acc_handle(Adapter->osdep.reg_handle) != DDI_FM_OK) {
816 		ddi_fm_service_impact(Adapter->dip, DDI_SERVICE_DEGRADED);
817 		Adapter->chip_state = E1000G_ERROR;
818 	}
819 
820 	return (ret_mp);
821 
822 rx_drop:
823 	/*
824 	 * Zero out the receive descriptors status
825 	 */
826 	current_desc->status = 0;
827 
828 	/* Sync the Rx descriptor DMA buffers */
829 	(void) ddi_dma_sync(rx_ring->rbd_dma_handle,
830 	    0, 0, DDI_DMA_SYNC_FORDEV);
831 
832 	if (current_desc == rx_ring->rbd_last)
833 		rx_ring->rbd_next = rx_ring->rbd_first;
834 	else
835 		rx_ring->rbd_next++;
836 
837 	last_desc = current_desc;
838 
839 	(p_rx_sw_packet_t)QUEUE_POP_HEAD(&rx_ring->recv_list);
840 
841 	QUEUE_PUSH_TAIL(&rx_ring->recv_list, &packet->Link);
842 	/*
843 	 * Reclaim all old buffers already allocated during
844 	 * Jumbo receives.....for incomplete reception
845 	 */
846 	if (rx_ring->rx_mblk != NULL) {
847 		freemsg(rx_ring->rx_mblk);
848 		rx_ring->rx_mblk = NULL;
849 		rx_ring->rx_mblk_tail = NULL;
850 		rx_ring->rx_mblk_len = 0;
851 	}
852 	/*
853 	 * Advance the E1000's Receive Queue #0 "Tail Pointer".
854 	 */
855 	E1000_WRITE_REG(hw, E1000_RDT,
856 	    (uint32_t)(last_desc - rx_ring->rbd_first));
857 
858 	if (e1000g_check_acc_handle(Adapter->osdep.reg_handle) != DDI_FM_OK) {
859 		ddi_fm_service_impact(Adapter->dip, DDI_SERVICE_DEGRADED);
860 		Adapter->chip_state = E1000G_ERROR;
861 	}
862 
863 	return (ret_mp);
864 }
865