xref: /titanic_41/usr/src/uts/common/io/igb/igb_rx.c (revision 8e50dcc9f00b393d43e6aa42b820bcbf1d3e1ce4)
1 /*
2  * CDDL HEADER START
3  *
4  * Copyright(c) 2007-2008 Intel Corporation. All rights reserved.
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at:
10  *	http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When using or redistributing this file, you may do so under the
15  * License only. No other modification of this header is permitted.
16  *
17  * If applicable, add the following below this CDDL HEADER, with the
18  * fields enclosed by brackets "[]" replaced with your own identifying
19  * information: Portions Copyright [yyyy] [name of copyright owner]
20  *
21  * CDDL HEADER END
22  */
23 
24 /*
25  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
26  * Use is subject to license terms of the CDDL.
27  */
28 
29 #pragma ident	"%Z%%M%	%I%	%E% SMI"
30 
31 #include "igb_sw.h"
32 
33 /* function prototypes */
34 static mblk_t *igb_rx_bind(igb_rx_ring_t *, uint32_t, uint32_t);
35 static mblk_t *igb_rx_copy(igb_rx_ring_t *, uint32_t, uint32_t);
36 static void igb_rx_assoc_hcksum(mblk_t *, uint32_t);
37 
38 #ifndef IGB_DEBUG
39 #pragma inline(igb_rx_assoc_hcksum)
40 #endif
41 
42 
43 /*
44  * igb_rx_recycle - the call-back function to reclaim rx buffer
45  *
46  * This function is called when an mp is freed by the user thru
47  * freeb call (Only for mp constructed through desballoc call).
48  * It returns back the freed buffer to the free list.
49  */
50 void
51 igb_rx_recycle(caddr_t arg)
52 {
53 	igb_rx_ring_t *rx_ring;
54 	rx_control_block_t *recycle_rcb;
55 	uint32_t free_index;
56 
57 	recycle_rcb = (rx_control_block_t *)(uintptr_t)arg;
58 	rx_ring = recycle_rcb->rx_ring;
59 
60 	if (recycle_rcb->state == RCB_FREE)
61 		return;
62 
63 	recycle_rcb->state = RCB_FREE;
64 
65 	ASSERT(recycle_rcb->mp == NULL);
66 
67 	/*
68 	 * Using the recycled data buffer to generate a new mblk
69 	 */
70 	recycle_rcb->mp = desballoc((unsigned char *)
71 	    (recycle_rcb->rx_buf.address - IPHDR_ALIGN_ROOM),
72 	    (recycle_rcb->rx_buf.size + IPHDR_ALIGN_ROOM),
73 	    0, &recycle_rcb->free_rtn);
74 	if (recycle_rcb->mp != NULL) {
75 		recycle_rcb->mp->b_rptr += IPHDR_ALIGN_ROOM;
76 		recycle_rcb->mp->b_wptr += IPHDR_ALIGN_ROOM;
77 	}
78 
79 	/*
80 	 * Put the recycled rx control block into free list
81 	 */
82 	mutex_enter(&rx_ring->recycle_lock);
83 
84 	free_index = rx_ring->rcb_tail;
85 	ASSERT(rx_ring->free_list[free_index] == NULL);
86 
87 	rx_ring->free_list[free_index] = recycle_rcb;
88 	rx_ring->rcb_tail = NEXT_INDEX(free_index, 1, rx_ring->free_list_size);
89 
90 	mutex_exit(&rx_ring->recycle_lock);
91 
92 	/*
93 	 * The atomic operation on the number of the available rx control
94 	 * blocks in the free list is used to make the recycling mutual
95 	 * exclusive with the receiving.
96 	 */
97 	atomic_inc_32(&rx_ring->rcb_free);
98 	ASSERT(rx_ring->rcb_free <= rx_ring->free_list_size);
99 }
100 
101 /*
102  * igb_rx_copy - Use copy to process the received packet
103  *
104  * This function will use bcopy to process the packet
105  * and send the copied packet upstream
106  */
107 static mblk_t *
108 igb_rx_copy(igb_rx_ring_t *rx_ring, uint32_t index, uint32_t pkt_len)
109 {
110 	rx_control_block_t *current_rcb;
111 	mblk_t *mp;
112 
113 	current_rcb = rx_ring->work_list[index];
114 
115 	DMA_SYNC(&current_rcb->rx_buf, DDI_DMA_SYNC_FORKERNEL);
116 
117 	/*
118 	 * Allocate buffer to receive this packet
119 	 */
120 	mp = allocb(pkt_len + IPHDR_ALIGN_ROOM, 0);
121 	if (mp == NULL) {
122 		igb_log(rx_ring->igb, "igb_rx_copy: allocate buffer failed");
123 		return (NULL);
124 	}
125 
126 	/*
127 	 * Copy the data received into the new cluster
128 	 */
129 	mp->b_rptr += IPHDR_ALIGN_ROOM;
130 	bcopy(current_rcb->rx_buf.address, mp->b_rptr, pkt_len);
131 	mp->b_wptr = mp->b_rptr + pkt_len;
132 
133 	return (mp);
134 }
135 
136 /*
137  * igb_rx_bind - Use existing DMA buffer to build mblk for receiving
138  *
139  * This function will use pre-bound DMA buffer to receive the packet
140  * and build mblk that will be sent upstream.
141  */
142 static mblk_t *
143 igb_rx_bind(igb_rx_ring_t *rx_ring, uint32_t index, uint32_t pkt_len)
144 {
145 	rx_control_block_t *current_rcb;
146 	rx_control_block_t *free_rcb;
147 	uint32_t free_index;
148 	mblk_t *mp;
149 
150 	/*
151 	 * If the free list is empty, we cannot proceed to send
152 	 * the current DMA buffer upstream. We'll have to return
153 	 * and use bcopy to process the packet.
154 	 */
155 	if (igb_atomic_reserve(&rx_ring->rcb_free, 1) < 0)
156 		return (NULL);
157 
158 	current_rcb = rx_ring->work_list[index];
159 	/*
160 	 * If the mp of the rx control block is NULL, try to do
161 	 * desballoc again.
162 	 */
163 	if (current_rcb->mp == NULL) {
164 		current_rcb->mp = desballoc((unsigned char *)
165 		    (current_rcb->rx_buf.address - IPHDR_ALIGN_ROOM),
166 		    (current_rcb->rx_buf.size + IPHDR_ALIGN_ROOM),
167 		    0, &current_rcb->free_rtn);
168 		/*
169 		 * If it is failed to built a mblk using the current
170 		 * DMA buffer, we have to return and use bcopy to
171 		 * process the packet.
172 		 */
173 		if (current_rcb->mp == NULL) {
174 			atomic_inc_32(&rx_ring->rcb_free);
175 			return (NULL);
176 		}
177 	}
178 	/*
179 	 * Sync up the data received
180 	 */
181 	DMA_SYNC(&current_rcb->rx_buf, DDI_DMA_SYNC_FORKERNEL);
182 
183 	mp = current_rcb->mp;
184 	current_rcb->mp = NULL;
185 	current_rcb->state = RCB_SENDUP;
186 
187 	mp->b_wptr = mp->b_rptr + pkt_len;
188 	mp->b_next = mp->b_cont = NULL;
189 
190 	/*
191 	 * Strip off one free rx control block from the free list
192 	 */
193 	free_index = rx_ring->rcb_head;
194 	free_rcb = rx_ring->free_list[free_index];
195 	ASSERT(free_rcb != NULL);
196 	rx_ring->free_list[free_index] = NULL;
197 	rx_ring->rcb_head = NEXT_INDEX(free_index, 1, rx_ring->free_list_size);
198 
199 	/*
200 	 * Put the rx control block to the work list
201 	 */
202 	rx_ring->work_list[index] = free_rcb;
203 
204 	return (mp);
205 }
206 
207 /*
208  * igb_rx_assoc_hcksum
209  *
210  * Check the rx hardware checksum status and associate the hcksum flags
211  */
212 static void
213 igb_rx_assoc_hcksum(mblk_t *mp, uint32_t status_error)
214 {
215 	uint32_t hcksum_flags = 0;
216 
217 	/* Ignore Checksum Indication */
218 	if (status_error & E1000_RXD_STAT_IXSM)
219 		return;
220 
221 	/*
222 	 * Check TCP/UDP checksum
223 	 */
224 	if (((status_error & E1000_RXD_STAT_TCPCS) ||
225 	    (status_error & E1000_RXD_STAT_UDPCS)) &&
226 	    !(status_error & E1000_RXDEXT_STATERR_TCPE))
227 		hcksum_flags |= HCK_FULLCKSUM | HCK_FULLCKSUM_OK;
228 
229 	/*
230 	 * Check IP Checksum
231 	 */
232 	if ((status_error & E1000_RXD_STAT_IPCS) &&
233 	    !(status_error & E1000_RXDEXT_STATERR_IPE))
234 		hcksum_flags |= HCK_IPV4_HDRCKSUM;
235 
236 	if (hcksum_flags != 0) {
237 		(void) hcksum_assoc(mp,
238 		    NULL, NULL, 0, 0, 0, 0, hcksum_flags, 0);
239 	}
240 }
241 
242 /*
243  * igb_rx - Receive the data of one ring
244  *
245  * This function goes throught h/w descriptor in one specified rx ring,
246  * receives the data if the descriptor status shows the data is ready.
247  * It returns a chain of mblks containing the received data, to be
248  * passed up to mac_rx().
249  */
250 mblk_t *
251 igb_rx(igb_rx_ring_t *rx_ring)
252 {
253 	union e1000_adv_rx_desc *current_rbd;
254 	rx_control_block_t *current_rcb;
255 	mblk_t *mp;
256 	mblk_t *mblk_head;
257 	mblk_t **mblk_tail;
258 	uint32_t rx_next;
259 	uint32_t rx_tail;
260 	uint32_t pkt_len;
261 	uint32_t status_error;
262 	uint32_t pkt_num;
263 	igb_t *igb = rx_ring->igb;
264 
265 	mblk_head = NULL;
266 	mblk_tail = &mblk_head;
267 
268 	/*
269 	 * Sync the receive descriptors before
270 	 * accepting the packets
271 	 */
272 	DMA_SYNC(&rx_ring->rbd_area, DDI_DMA_SYNC_FORKERNEL);
273 
274 	/*
275 	 * Get the start point of rx bd ring which should be examined
276 	 * during this cycle.
277 	 */
278 	rx_next = rx_ring->rbd_next;
279 
280 	current_rbd = &rx_ring->rbd_ring[rx_next];
281 	pkt_num = 0;
282 	status_error = current_rbd->wb.upper.status_error;
283 	while (status_error & E1000_RXD_STAT_DD) {
284 		/*
285 		 * If hardware has found the errors, but the error
286 		 * is hardware checksum error, here does not discard the
287 		 * packet, and let upper layer compute the checksum;
288 		 * Otherwise discard the packet.
289 		 */
290 		if ((status_error & E1000_RXDEXT_ERR_FRAME_ERR_MASK) ||
291 		    !(status_error & E1000_RXD_STAT_EOP)) {
292 			IGB_DEBUG_STAT(rx_ring->stat_frame_error);
293 			goto rx_discard;
294 		}
295 
296 		IGB_DEBUG_STAT_COND(rx_ring->stat_cksum_error,
297 		    (status_error & E1000_RXDEXT_STATERR_TCPE) ||
298 		    (status_error & E1000_RXDEXT_STATERR_IPE));
299 
300 		pkt_len = current_rbd->wb.upper.length;
301 		mp = NULL;
302 		/*
303 		 * For packets with length more than the copy threshold,
304 		 * we'll firstly try to use the existed DMA buffer to built
305 		 * a mblk and send the mblk upstream.
306 		 *
307 		 * If the first method fails, or the packet length is less
308 		 * than the copy threshold, we'll allocate a new mblk and
309 		 * copy the packet data to the mblk.
310 		 */
311 		if (pkt_len > rx_ring->copy_thresh)
312 			mp = igb_rx_bind(rx_ring, rx_next, pkt_len);
313 
314 		if (mp == NULL)
315 			mp = igb_rx_copy(rx_ring, rx_next, pkt_len);
316 
317 		if (mp != NULL) {
318 			/*
319 			 * Check h/w checksum offload status
320 			 */
321 			if (igb->rx_hcksum_enable)
322 				igb_rx_assoc_hcksum(mp, status_error);
323 
324 			*mblk_tail = mp;
325 			mblk_tail = &mp->b_next;
326 		}
327 
328 rx_discard:
329 		/*
330 		 * Reset rx descriptor read bits
331 		 */
332 		current_rcb = rx_ring->work_list[rx_next];
333 		current_rbd->read.pkt_addr = current_rcb->rx_buf.dma_address;
334 		current_rbd->read.hdr_addr = 0;
335 
336 		rx_next = NEXT_INDEX(rx_next, 1, rx_ring->ring_size);
337 
338 		/*
339 		 * The receive function is in interrupt context, so here
340 		 * limit_per_intr is used to avoid doing receiving too long
341 		 * per interrupt.
342 		 */
343 		if (++pkt_num > rx_ring->limit_per_intr) {
344 			IGB_DEBUG_STAT(rx_ring->stat_exceed_pkt);
345 			break;
346 		}
347 
348 		current_rbd = &rx_ring->rbd_ring[rx_next];
349 		status_error = current_rbd->wb.upper.status_error;
350 	}
351 
352 	DMA_SYNC(&rx_ring->rbd_area, DDI_DMA_SYNC_FORDEV);
353 
354 	rx_ring->rbd_next = rx_next;
355 
356 	/*
357 	 * Update the h/w tail accordingly
358 	 */
359 	rx_tail = PREV_INDEX(rx_next, 1, rx_ring->ring_size);
360 
361 	E1000_WRITE_REG(&igb->hw, E1000_RDT(rx_ring->index), rx_tail);
362 
363 	return (mblk_head);
364 }
365