xref: /illumos-gate/usr/src/uts/common/io/bge/bge_recv2.c (revision fcdb3229a31dd4ff700c69238814e326aad49098)
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 "bge_impl.h"
28 
29 #define	U32TOPTR(x)	((void *)(uintptr_t)(uint32_t)(x))
30 #define	PTRTOU32(x)	((uint32_t)(uintptr_t)(void *)(x))
31 
32 /*
33  * ========== RX side routines ==========
34  */
35 
36 #define	BGE_DBG		BGE_DBG_RECV	/* debug flag for this code	*/
37 
38 static void bge_refill(bge_t *bgep, buff_ring_t *brp, sw_rbd_t *srbdp);
39 
40 /*
41  * Return the specified buffer (srbdp) to the ring it came from (brp).
42  *
43  * Note:
44  *	If the driver is compiled with only one buffer ring *and* one
45  *	return ring, then the buffers must be returned in sequence.
46  *	In this case, we don't have to consider anything about the
47  *	buffer at all; we can simply advance the cyclic counter.  And
48  *	we don't even need the refill mutex <rf_lock>, as the caller
49  *	will already be holding the (one-and-only) <rx_lock>.
50  *
51  *	If the driver supports multiple buffer rings, but only one
52  *	return ring, the same still applies (to each buffer ring
53  *	separately).
54  */
55 static void
bge_refill(bge_t * bgep,buff_ring_t * brp,sw_rbd_t * srbdp)56 bge_refill(bge_t *bgep, buff_ring_t *brp, sw_rbd_t *srbdp)
57 {
58 	uint64_t slot;
59 
60 	_NOTE(ARGUNUSED(srbdp))
61 
62 	slot = brp->rf_next;
63 	brp->rf_next = NEXT(slot, brp->desc.nslots);
64 	bge_mbx_put(bgep, brp->chip_mbx_reg, slot);
65 }
66 
67 static mblk_t *bge_receive_packet(bge_t *bgep, bge_rbd_t *hw_rbd_p,
68     recv_ring_t *rrp);
69 
70 static mblk_t *
bge_receive_packet(bge_t * bgep,bge_rbd_t * hw_rbd_p,recv_ring_t * rrp)71 bge_receive_packet(bge_t *bgep, bge_rbd_t *hw_rbd_p, recv_ring_t *rrp)
72 {
73 	bge_rbd_t hw_rbd;
74 	buff_ring_t *brp;
75 	sw_rbd_t *srbdp;
76 	uchar_t *dp;
77 	mblk_t *mp;
78 	uint_t len;
79 	uint_t minsize;
80 	uint_t maxsize;
81 	uint32_t pflags;
82 
83 	mp = NULL;
84 	hw_rbd = *hw_rbd_p;
85 
86 	switch (hw_rbd.flags & (RBD_FLAG_MINI_RING|RBD_FLAG_JUMBO_RING)) {
87 	case RBD_FLAG_MINI_RING|RBD_FLAG_JUMBO_RING:
88 	default:
89 		/* error, this shouldn't happen */
90 		BGE_PKTDUMP((bgep, &hw_rbd, NULL, "bad ring flags!"));
91 		goto error;
92 
93 	case RBD_FLAG_JUMBO_RING:
94 		brp = &bgep->buff[BGE_JUMBO_BUFF_RING];
95 		break;
96 
97 #if	(BGE_BUFF_RINGS_USED > 2)
98 	case RBD_FLAG_MINI_RING:
99 		brp = &bgep->buff[BGE_MINI_BUFF_RING];
100 		break;
101 #endif	/* BGE_BUFF_RINGS_USED > 2 */
102 
103 	case 0:
104 		brp = &bgep->buff[BGE_STD_BUFF_RING];
105 		break;
106 	}
107 
108 	if (hw_rbd.index >= brp->desc.nslots) {
109 		/* error, this shouldn't happen */
110 		BGE_PKTDUMP((bgep, &hw_rbd, NULL, "bad ring index!"));
111 		goto error;
112 	}
113 
114 	srbdp = &brp->sw_rbds[hw_rbd.index];
115 	if (hw_rbd.opaque != srbdp->pbuf.token) {
116 		/* bogus, drop the packet */
117 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "bad ring token"));
118 		goto refill;
119 	}
120 
121 	if ((hw_rbd.flags & RBD_FLAG_PACKET_END) == 0) {
122 		/* bogus, drop the packet */
123 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "unterminated packet"));
124 		goto refill;
125 	}
126 
127 	if (hw_rbd.flags & RBD_FLAG_FRAME_HAS_ERROR) {
128 		/* bogus, drop the packet */
129 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "errored packet"));
130 		goto refill;
131 	}
132 
133 	len = hw_rbd.len;
134 
135 #ifdef BGE_IPMI_ASF
136 	/*
137 	 * When IPMI/ASF is enabled, VLAN tag must be stripped.
138 	 */
139 	if (bgep->asf_enabled && (hw_rbd.flags & RBD_FLAG_VLAN_TAG))
140 		maxsize = bgep->chipid.ethmax_size + ETHERFCSL;
141 	else
142 #endif
143 		/*
144 		 * H/W will not strip the VLAN tag from incoming packet
145 		 * now, as RECEIVE_MODE_KEEP_VLAN_TAG bit is set in
146 		 * RECEIVE_MAC_MODE_REG register.
147 		 */
148 		maxsize = bgep->chipid.ethmax_size + VLAN_TAGSZ + ETHERFCSL;
149 	if (len > maxsize) {
150 		/* bogus, drop the packet */
151 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "oversize packet"));
152 		goto refill;
153 	}
154 
155 #ifdef BGE_IPMI_ASF
156 	if (bgep->asf_enabled && (hw_rbd.flags & RBD_FLAG_VLAN_TAG))
157 		minsize = ETHERMIN + ETHERFCSL - VLAN_TAGSZ;
158 	else
159 #endif
160 		minsize = ETHERMIN + ETHERFCSL;
161 	if (len < minsize) {
162 		/* bogus, drop the packet */
163 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "undersize packet"));
164 		goto refill;
165 	}
166 
167 	/*
168 	 * Packet looks good; get a buffer to copy it into.
169 	 * We want to leave some space at the front of the allocated
170 	 * buffer in case any upstream modules want to prepend some
171 	 * sort of header.  This also has the side-effect of making
172 	 * the packet *contents* 4-byte aligned, as required by NCA!
173 	 */
174 #ifdef BGE_IPMI_ASF
175 	if (bgep->asf_enabled && (hw_rbd.flags & RBD_FLAG_VLAN_TAG)) {
176 		mp = allocb(BGE_HEADROOM + len + VLAN_TAGSZ, 0);
177 	} else {
178 #endif
179 
180 		mp = allocb(BGE_HEADROOM + len, 0);
181 #ifdef BGE_IPMI_ASF
182 	}
183 #endif
184 	if (mp == NULL) {
185 		/* Nothing to do but drop the packet */
186 		goto refill;
187 	}
188 
189 	/*
190 	 * Sync the data and copy it to the STREAMS buffer.
191 	 */
192 	DMA_SYNC(srbdp->pbuf, DDI_DMA_SYNC_FORKERNEL);
193 	if (bge_check_dma_handle(bgep, srbdp->pbuf.dma_hdl) != DDI_FM_OK) {
194 		bgep->bge_dma_error = B_TRUE;
195 		bgep->bge_chip_state = BGE_CHIP_ERROR;
196 		return (NULL);
197 	}
198 #ifdef BGE_IPMI_ASF
199 	if (bgep->asf_enabled && (hw_rbd.flags & RBD_FLAG_VLAN_TAG)) {
200 		/*
201 		 * As VLAN tag has been stripped from incoming packet in ASF
202 		 * scenario, we insert it into this packet again.
203 		 */
204 		struct ether_vlan_header *ehp;
205 		mp->b_rptr = dp = mp->b_rptr + BGE_HEADROOM - VLAN_TAGSZ;
206 		bcopy(DMA_VPTR(srbdp->pbuf), dp, 2 * ETHERADDRL);
207 		ehp = (void *)dp;
208 		ehp->ether_tpid = ntohs(ETHERTYPE_VLAN);
209 		ehp->ether_tci = ntohs(hw_rbd.vlan_tci);
210 		bcopy(((uchar_t *)(DMA_VPTR(srbdp->pbuf))) + 2 * ETHERADDRL,
211 		    dp + 2 * ETHERADDRL + VLAN_TAGSZ,
212 		    len - 2 * ETHERADDRL);
213 	} else {
214 #endif
215 		mp->b_rptr = dp = mp->b_rptr + BGE_HEADROOM;
216 		bcopy(DMA_VPTR(srbdp->pbuf), dp, len);
217 #ifdef BGE_IPMI_ASF
218 	}
219 
220 	if (bgep->asf_enabled && (hw_rbd.flags & RBD_FLAG_VLAN_TAG)) {
221 		mp->b_wptr = dp + len + VLAN_TAGSZ - ETHERFCSL;
222 	} else
223 #endif
224 		mp->b_wptr = dp + len - ETHERFCSL;
225 
226 	/*
227 	 * Special check for one specific type of data corruption;
228 	 * in a good packet, the first 8 bytes are *very* unlikely
229 	 * to be the same as the second 8 bytes ... but we let the
230 	 * packet through just in case.
231 	 */
232 	if (bcmp(dp, dp+8, 8) == 0)
233 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "stuttered packet?"));
234 
235 	pflags = 0;
236 	if (hw_rbd.flags & RBD_FLAG_TCP_UDP_CHECKSUM)
237 		pflags |= HCK_FULLCKSUM;
238 	if (hw_rbd.flags & RBD_FLAG_IP_CHECKSUM)
239 		pflags |= HCK_IPV4_HDRCKSUM_OK;
240 	if (pflags != 0)
241 		mac_hcksum_set(mp, 0, 0, 0, hw_rbd.tcp_udp_cksum, pflags);
242 
243 	/* Update per-ring rx statistics */
244 	rrp->rx_pkts++;
245 	rrp->rx_bytes += len;
246 
247 refill:
248 	/*
249 	 * Replace the buffer in the ring it came from ...
250 	 */
251 	bge_refill(bgep, brp, srbdp);
252 	return (mp);
253 
254 error:
255 	/*
256 	 * We come here if the integrity of the ring descriptors
257 	 * (rather than merely packet data) appears corrupted.
258 	 * The factotum will attempt to reset-and-recover.
259 	 */
260 	bgep->bge_chip_state = BGE_CHIP_ERROR;
261 	bge_fm_ereport(bgep, DDI_FM_DEVICE_INVAL_STATE);
262 	return (NULL);
263 }
264 
265 /*
266  * Accept the packets received in the specified ring up to
267  * (but not including) the producer index in the status block.
268  *
269  * Returns a chain of mblks containing the received data, to be
270  * passed up to gld_recv() (we can't call gld_recv() from here,
271  * 'cos we're holding the per-ring receive lock at this point).
272  *
273  * This function must advance (rrp->rx_next) and write it back to
274  * the chip to indicate the packets it has accepted from the ring.
275  */
276 static mblk_t *bge_receive_ring(bge_t *bgep, recv_ring_t *rrp);
277 
278 static mblk_t *
bge_receive_ring(bge_t * bgep,recv_ring_t * rrp)279 bge_receive_ring(bge_t *bgep, recv_ring_t *rrp)
280 {
281 	bge_rbd_t *hw_rbd_p;
282 	uint64_t slot;
283 	mblk_t *head;
284 	mblk_t **tail;
285 	mblk_t *mp;
286 	int recv_cnt = 0;
287 
288 	ASSERT(mutex_owned(rrp->rx_lock));
289 
290 	/*
291 	 * Sync (all) the receive ring descriptors
292 	 * before accepting the packets they describe
293 	 */
294 	DMA_SYNC(rrp->desc, DDI_DMA_SYNC_FORKERNEL);
295 	if (*rrp->prod_index_p >= rrp->desc.nslots) {
296 		bgep->bge_chip_state = BGE_CHIP_ERROR;
297 		bge_fm_ereport(bgep, DDI_FM_DEVICE_INVAL_STATE);
298 		return (NULL);
299 	}
300 	if (bge_check_dma_handle(bgep, rrp->desc.dma_hdl) != DDI_FM_OK) {
301 		rrp->rx_next = *rrp->prod_index_p;
302 		bge_mbx_put(bgep, rrp->chip_mbx_reg, rrp->rx_next);
303 		bgep->bge_dma_error = B_TRUE;
304 		bgep->bge_chip_state = BGE_CHIP_ERROR;
305 		return (NULL);
306 	}
307 
308 	hw_rbd_p = DMA_VPTR(rrp->desc);
309 	head = NULL;
310 	tail = &head;
311 	slot = rrp->rx_next;
312 
313 	while ((slot != *rrp->prod_index_p) && /* Note: volatile	*/
314 	    (recv_cnt < BGE_MAXPKT_RCVED)) {
315 		if ((mp = bge_receive_packet(bgep, &hw_rbd_p[slot], rrp))
316 		    != NULL) {
317 			*tail = mp;
318 			tail = &mp->b_next;
319 			recv_cnt++;
320 		}
321 		rrp->rx_next = slot = NEXT(slot, rrp->desc.nslots);
322 	}
323 
324 	bge_mbx_put(bgep, rrp->chip_mbx_reg, rrp->rx_next);
325 	if (bge_check_acc_handle(bgep, bgep->io_handle) != DDI_FM_OK)
326 		bgep->bge_chip_state = BGE_CHIP_ERROR;
327 	return (head);
328 }
329 
330 /*
331  * XXX: Poll a particular ring. The implementation is incomplete.
332  * Once the ring interrupts are disabled, we need to do bge_recyle()
333  * for the ring as well and re enable the ring interrupt automatically
334  * if the poll doesn't find any packets in the ring. We need to
335  * have MSI-X interrupts support for this.
336  *
337  * The basic poll policy is that rings that are dealing with explicit
338  * flows (like TCP or some service) and are marked as such should
339  * have their own MSI-X interrupt per ring. bge_intr() should leave
340  * that interrupt disabled after an upcall. The ring is in poll mode.
341  * When a poll thread comes down and finds nothing, the MSI-X interrupt
342  * is automatically enabled. Squeue needs to deal with the race of
343  * a new interrupt firing and reaching before poll thread returns.
344  */
345 mblk_t *
bge_poll_ring(void * arg,int bytes_to_pickup)346 bge_poll_ring(void *arg, int bytes_to_pickup)
347 {
348 	recv_ring_t *rrp = arg;
349 	bge_t *bgep = rrp->bgep;
350 	bge_rbd_t *hw_rbd_p;
351 	uint64_t slot;
352 	mblk_t *head;
353 	mblk_t **tail;
354 	mblk_t *mp;
355 	size_t sz = 0;
356 
357 	mutex_enter(rrp->rx_lock);
358 
359 	/*
360 	 * Sync (all) the receive ring descriptors
361 	 * before accepting the packets they describe
362 	 */
363 	DMA_SYNC(rrp->desc, DDI_DMA_SYNC_FORKERNEL);
364 	if (*rrp->prod_index_p >= rrp->desc.nslots) {
365 		bgep->bge_chip_state = BGE_CHIP_ERROR;
366 		bge_fm_ereport(bgep, DDI_FM_DEVICE_INVAL_STATE);
367 		mutex_exit(rrp->rx_lock);
368 		return (NULL);
369 	}
370 	if (bge_check_dma_handle(bgep, rrp->desc.dma_hdl) != DDI_FM_OK) {
371 		rrp->rx_next = *rrp->prod_index_p;
372 		bge_mbx_put(bgep, rrp->chip_mbx_reg, rrp->rx_next);
373 		bgep->bge_dma_error = B_TRUE;
374 		bgep->bge_chip_state = BGE_CHIP_ERROR;
375 		mutex_exit(rrp->rx_lock);
376 		return (NULL);
377 	}
378 
379 	hw_rbd_p = DMA_VPTR(rrp->desc);
380 	head = NULL;
381 	tail = &head;
382 	slot = rrp->rx_next;
383 
384 	/* Note: volatile */
385 	while ((slot != *rrp->prod_index_p) && (sz <= bytes_to_pickup)) {
386 		if ((mp = bge_receive_packet(bgep, &hw_rbd_p[slot], rrp))
387 		    != NULL) {
388 			*tail = mp;
389 			sz += msgdsize(mp);
390 			tail = &mp->b_next;
391 		}
392 		rrp->rx_next = slot = NEXT(slot, rrp->desc.nslots);
393 	}
394 
395 	bge_mbx_put(bgep, rrp->chip_mbx_reg, rrp->rx_next);
396 	if (bge_check_acc_handle(bgep, bgep->io_handle) != DDI_FM_OK)
397 		bgep->bge_chip_state = BGE_CHIP_ERROR;
398 	mutex_exit(rrp->rx_lock);
399 	return (head);
400 }
401 
402 /*
403  * Receive all packets in all rings.
404  */
405 void bge_receive(bge_t *bgep, bge_status_t *bsp);
406 
407 void
bge_receive(bge_t * bgep,bge_status_t * bsp)408 bge_receive(bge_t *bgep, bge_status_t *bsp)
409 {
410 	recv_ring_t *rrp;
411 	uint64_t index;
412 	mblk_t *mp;
413 
414 	for (index = 0; index < bgep->chipid.rx_rings; index++) {
415 		/*
416 		 * Start from the first ring.
417 		 */
418 		rrp = &bgep->recv[index];
419 
420 		/*
421 		 * For each ring, (rrp->prod_index_p) points to the
422 		 * proper index within the status block (which has
423 		 * already been sync'd by the caller)
424 		 */
425 		ASSERT(rrp->prod_index_p == RECV_INDEX_P(bsp, index));
426 
427 		if (*rrp->prod_index_p == rrp->rx_next || rrp->poll_flag)
428 			continue;		/* no packets		*/
429 		if (mutex_tryenter(rrp->rx_lock) == 0)
430 			continue;		/* already in process	*/
431 		mp = bge_receive_ring(bgep, rrp);
432 		mutex_exit(rrp->rx_lock);
433 
434 		if (mp != NULL)
435 			mac_rx_ring(bgep->mh, rrp->ring_handle, mp,
436 			    rrp->ring_gen_num);
437 	}
438 }
439