xref: /linux/drivers/infiniband/hw/hfi1/ud.c (revision f3a8b6645dc2e60d11f20c1c23afd964ff4e55ae)
1 /*
2  * Copyright(c) 2015, 2016 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <linux/net.h>
49 #include <rdma/ib_smi.h>
50 
51 #include "hfi.h"
52 #include "mad.h"
53 #include "verbs_txreq.h"
54 #include "qp.h"
55 
56 /**
57  * ud_loopback - handle send on loopback QPs
58  * @sqp: the sending QP
59  * @swqe: the send work request
60  *
61  * This is called from hfi1_make_ud_req() to forward a WQE addressed
62  * to the same HFI.
63  * Note that the receive interrupt handler may be calling hfi1_ud_rcv()
64  * while this is being called.
65  */
66 static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
67 {
68 	struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
69 	struct hfi1_pportdata *ppd;
70 	struct rvt_qp *qp;
71 	struct ib_ah_attr *ah_attr;
72 	unsigned long flags;
73 	struct rvt_sge_state ssge;
74 	struct rvt_sge *sge;
75 	struct ib_wc wc;
76 	u32 length;
77 	enum ib_qp_type sqptype, dqptype;
78 
79 	rcu_read_lock();
80 
81 	qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
82 			    swqe->ud_wr.remote_qpn);
83 	if (!qp) {
84 		ibp->rvp.n_pkt_drops++;
85 		rcu_read_unlock();
86 		return;
87 	}
88 
89 	sqptype = sqp->ibqp.qp_type == IB_QPT_GSI ?
90 			IB_QPT_UD : sqp->ibqp.qp_type;
91 	dqptype = qp->ibqp.qp_type == IB_QPT_GSI ?
92 			IB_QPT_UD : qp->ibqp.qp_type;
93 
94 	if (dqptype != sqptype ||
95 	    !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
96 		ibp->rvp.n_pkt_drops++;
97 		goto drop;
98 	}
99 
100 	ah_attr = &ibah_to_rvtah(swqe->ud_wr.ah)->attr;
101 	ppd = ppd_from_ibp(ibp);
102 
103 	if (qp->ibqp.qp_num > 1) {
104 		u16 pkey;
105 		u16 slid;
106 		u8 sc5 = ibp->sl_to_sc[ah_attr->sl];
107 
108 		pkey = hfi1_get_pkey(ibp, sqp->s_pkey_index);
109 		slid = ppd->lid | (ah_attr->src_path_bits &
110 				   ((1 << ppd->lmc) - 1));
111 		if (unlikely(ingress_pkey_check(ppd, pkey, sc5,
112 						qp->s_pkey_index, slid))) {
113 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY, pkey,
114 				       ah_attr->sl,
115 				       sqp->ibqp.qp_num, qp->ibqp.qp_num,
116 				       slid, ah_attr->dlid);
117 			goto drop;
118 		}
119 	}
120 
121 	/*
122 	 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
123 	 * Qkeys with the high order bit set mean use the
124 	 * qkey from the QP context instead of the WR (see 10.2.5).
125 	 */
126 	if (qp->ibqp.qp_num) {
127 		u32 qkey;
128 
129 		qkey = (int)swqe->ud_wr.remote_qkey < 0 ?
130 			sqp->qkey : swqe->ud_wr.remote_qkey;
131 		if (unlikely(qkey != qp->qkey)) {
132 			u16 lid;
133 
134 			lid = ppd->lid | (ah_attr->src_path_bits &
135 					  ((1 << ppd->lmc) - 1));
136 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_Q_KEY, qkey,
137 				       ah_attr->sl,
138 				       sqp->ibqp.qp_num, qp->ibqp.qp_num,
139 				       lid,
140 				       ah_attr->dlid);
141 			goto drop;
142 		}
143 	}
144 
145 	/*
146 	 * A GRH is expected to precede the data even if not
147 	 * present on the wire.
148 	 */
149 	length = swqe->length;
150 	memset(&wc, 0, sizeof(wc));
151 	wc.byte_len = length + sizeof(struct ib_grh);
152 
153 	if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
154 		wc.wc_flags = IB_WC_WITH_IMM;
155 		wc.ex.imm_data = swqe->wr.ex.imm_data;
156 	}
157 
158 	spin_lock_irqsave(&qp->r_lock, flags);
159 
160 	/*
161 	 * Get the next work request entry to find where to put the data.
162 	 */
163 	if (qp->r_flags & RVT_R_REUSE_SGE) {
164 		qp->r_flags &= ~RVT_R_REUSE_SGE;
165 	} else {
166 		int ret;
167 
168 		ret = hfi1_rvt_get_rwqe(qp, 0);
169 		if (ret < 0) {
170 			hfi1_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
171 			goto bail_unlock;
172 		}
173 		if (!ret) {
174 			if (qp->ibqp.qp_num == 0)
175 				ibp->rvp.n_vl15_dropped++;
176 			goto bail_unlock;
177 		}
178 	}
179 	/* Silently drop packets which are too big. */
180 	if (unlikely(wc.byte_len > qp->r_len)) {
181 		qp->r_flags |= RVT_R_REUSE_SGE;
182 		ibp->rvp.n_pkt_drops++;
183 		goto bail_unlock;
184 	}
185 
186 	if (ah_attr->ah_flags & IB_AH_GRH) {
187 		struct ib_grh grh;
188 		struct ib_global_route grd = ah_attr->grh;
189 
190 		hfi1_make_grh(ibp, &grh, &grd, 0, 0);
191 		hfi1_copy_sge(&qp->r_sge, &grh,
192 			      sizeof(grh), 1, 0);
193 		wc.wc_flags |= IB_WC_GRH;
194 	} else {
195 		hfi1_skip_sge(&qp->r_sge, sizeof(struct ib_grh), 1);
196 	}
197 	ssge.sg_list = swqe->sg_list + 1;
198 	ssge.sge = *swqe->sg_list;
199 	ssge.num_sge = swqe->wr.num_sge;
200 	sge = &ssge.sge;
201 	while (length) {
202 		u32 len = sge->length;
203 
204 		if (len > length)
205 			len = length;
206 		if (len > sge->sge_length)
207 			len = sge->sge_length;
208 		WARN_ON_ONCE(len == 0);
209 		hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, 1, 0);
210 		sge->vaddr += len;
211 		sge->length -= len;
212 		sge->sge_length -= len;
213 		if (sge->sge_length == 0) {
214 			if (--ssge.num_sge)
215 				*sge = *ssge.sg_list++;
216 		} else if (sge->length == 0 && sge->mr->lkey) {
217 			if (++sge->n >= RVT_SEGSZ) {
218 				if (++sge->m >= sge->mr->mapsz)
219 					break;
220 				sge->n = 0;
221 			}
222 			sge->vaddr =
223 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
224 			sge->length =
225 				sge->mr->map[sge->m]->segs[sge->n].length;
226 		}
227 		length -= len;
228 	}
229 	rvt_put_ss(&qp->r_sge);
230 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
231 		goto bail_unlock;
232 	wc.wr_id = qp->r_wr_id;
233 	wc.status = IB_WC_SUCCESS;
234 	wc.opcode = IB_WC_RECV;
235 	wc.qp = &qp->ibqp;
236 	wc.src_qp = sqp->ibqp.qp_num;
237 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI) {
238 		if (sqp->ibqp.qp_type == IB_QPT_GSI ||
239 		    sqp->ibqp.qp_type == IB_QPT_SMI)
240 			wc.pkey_index = swqe->ud_wr.pkey_index;
241 		else
242 			wc.pkey_index = sqp->s_pkey_index;
243 	} else {
244 		wc.pkey_index = 0;
245 	}
246 	wc.slid = ppd->lid | (ah_attr->src_path_bits & ((1 << ppd->lmc) - 1));
247 	/* Check for loopback when the port lid is not set */
248 	if (wc.slid == 0 && sqp->ibqp.qp_type == IB_QPT_GSI)
249 		wc.slid = be16_to_cpu(IB_LID_PERMISSIVE);
250 	wc.sl = ah_attr->sl;
251 	wc.dlid_path_bits = ah_attr->dlid & ((1 << ppd->lmc) - 1);
252 	wc.port_num = qp->port_num;
253 	/* Signal completion event if the solicited bit is set. */
254 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
255 		     swqe->wr.send_flags & IB_SEND_SOLICITED);
256 	ibp->rvp.n_loop_pkts++;
257 bail_unlock:
258 	spin_unlock_irqrestore(&qp->r_lock, flags);
259 drop:
260 	rcu_read_unlock();
261 }
262 
263 /**
264  * hfi1_make_ud_req - construct a UD request packet
265  * @qp: the QP
266  *
267  * Assume s_lock is held.
268  *
269  * Return 1 if constructed; otherwise, return 0.
270  */
271 int hfi1_make_ud_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
272 {
273 	struct hfi1_qp_priv *priv = qp->priv;
274 	struct ib_other_headers *ohdr;
275 	struct ib_ah_attr *ah_attr;
276 	struct hfi1_pportdata *ppd;
277 	struct hfi1_ibport *ibp;
278 	struct rvt_swqe *wqe;
279 	u32 nwords;
280 	u32 extra_bytes;
281 	u32 bth0;
282 	u16 lrh0;
283 	u16 lid;
284 	int next_cur;
285 	u8 sc5;
286 
287 	ps->s_txreq = get_txreq(ps->dev, qp);
288 	if (IS_ERR(ps->s_txreq))
289 		goto bail_no_tx;
290 
291 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
292 		if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
293 			goto bail;
294 		/* We are in the error state, flush the work request. */
295 		smp_read_barrier_depends(); /* see post_one_send */
296 		if (qp->s_last == ACCESS_ONCE(qp->s_head))
297 			goto bail;
298 		/* If DMAs are in progress, we can't flush immediately. */
299 		if (iowait_sdma_pending(&priv->s_iowait)) {
300 			qp->s_flags |= RVT_S_WAIT_DMA;
301 			goto bail;
302 		}
303 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
304 		hfi1_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
305 		goto done_free_tx;
306 	}
307 
308 	/* see post_one_send() */
309 	smp_read_barrier_depends();
310 	if (qp->s_cur == ACCESS_ONCE(qp->s_head))
311 		goto bail;
312 
313 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
314 	next_cur = qp->s_cur + 1;
315 	if (next_cur >= qp->s_size)
316 		next_cur = 0;
317 
318 	/* Construct the header. */
319 	ibp = to_iport(qp->ibqp.device, qp->port_num);
320 	ppd = ppd_from_ibp(ibp);
321 	ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr;
322 	if (ah_attr->dlid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
323 	    ah_attr->dlid == be16_to_cpu(IB_LID_PERMISSIVE)) {
324 		lid = ah_attr->dlid & ~((1 << ppd->lmc) - 1);
325 		if (unlikely(!loopback &&
326 			     (lid == ppd->lid ||
327 			      (lid == be16_to_cpu(IB_LID_PERMISSIVE) &&
328 			      qp->ibqp.qp_type == IB_QPT_GSI)))) {
329 			unsigned long tflags = ps->flags;
330 			/*
331 			 * If DMAs are in progress, we can't generate
332 			 * a completion for the loopback packet since
333 			 * it would be out of order.
334 			 * Instead of waiting, we could queue a
335 			 * zero length descriptor so we get a callback.
336 			 */
337 			if (iowait_sdma_pending(&priv->s_iowait)) {
338 				qp->s_flags |= RVT_S_WAIT_DMA;
339 				goto bail;
340 			}
341 			qp->s_cur = next_cur;
342 			spin_unlock_irqrestore(&qp->s_lock, tflags);
343 			ud_loopback(qp, wqe);
344 			spin_lock_irqsave(&qp->s_lock, tflags);
345 			ps->flags = tflags;
346 			hfi1_send_complete(qp, wqe, IB_WC_SUCCESS);
347 			goto done_free_tx;
348 		}
349 	}
350 
351 	qp->s_cur = next_cur;
352 	extra_bytes = -wqe->length & 3;
353 	nwords = (wqe->length + extra_bytes) >> 2;
354 
355 	/* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
356 	qp->s_hdrwords = 7;
357 	qp->s_cur_size = wqe->length;
358 	qp->s_cur_sge = &qp->s_sge;
359 	qp->s_srate = ah_attr->static_rate;
360 	qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
361 	qp->s_wqe = wqe;
362 	qp->s_sge.sge = wqe->sg_list[0];
363 	qp->s_sge.sg_list = wqe->sg_list + 1;
364 	qp->s_sge.num_sge = wqe->wr.num_sge;
365 	qp->s_sge.total_len = wqe->length;
366 
367 	if (ah_attr->ah_flags & IB_AH_GRH) {
368 		/* Header size in 32-bit words. */
369 		qp->s_hdrwords += hfi1_make_grh(ibp,
370 						&ps->s_txreq->phdr.hdr.u.l.grh,
371 						&ah_attr->grh,
372 						qp->s_hdrwords, nwords);
373 		lrh0 = HFI1_LRH_GRH;
374 		ohdr = &ps->s_txreq->phdr.hdr.u.l.oth;
375 		/*
376 		 * Don't worry about sending to locally attached multicast
377 		 * QPs.  It is unspecified by the spec. what happens.
378 		 */
379 	} else {
380 		/* Header size in 32-bit words. */
381 		lrh0 = HFI1_LRH_BTH;
382 		ohdr = &ps->s_txreq->phdr.hdr.u.oth;
383 	}
384 	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
385 		qp->s_hdrwords++;
386 		ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
387 		bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
388 	} else {
389 		bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
390 	}
391 	sc5 = ibp->sl_to_sc[ah_attr->sl];
392 	lrh0 |= (ah_attr->sl & 0xf) << 4;
393 	if (qp->ibqp.qp_type == IB_QPT_SMI) {
394 		lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
395 		priv->s_sc = 0xf;
396 	} else {
397 		lrh0 |= (sc5 & 0xf) << 12;
398 		priv->s_sc = sc5;
399 	}
400 	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
401 	ps->s_txreq->sde = priv->s_sde;
402 	priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
403 	ps->s_txreq->psc = priv->s_sendcontext;
404 	ps->s_txreq->phdr.hdr.lrh[0] = cpu_to_be16(lrh0);
405 	ps->s_txreq->phdr.hdr.lrh[1] = cpu_to_be16(ah_attr->dlid);
406 	ps->s_txreq->phdr.hdr.lrh[2] =
407 		cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
408 	if (ah_attr->dlid == be16_to_cpu(IB_LID_PERMISSIVE)) {
409 		ps->s_txreq->phdr.hdr.lrh[3] = IB_LID_PERMISSIVE;
410 	} else {
411 		lid = ppd->lid;
412 		if (lid) {
413 			lid |= ah_attr->src_path_bits & ((1 << ppd->lmc) - 1);
414 			ps->s_txreq->phdr.hdr.lrh[3] = cpu_to_be16(lid);
415 		} else {
416 			ps->s_txreq->phdr.hdr.lrh[3] = IB_LID_PERMISSIVE;
417 		}
418 	}
419 	if (wqe->wr.send_flags & IB_SEND_SOLICITED)
420 		bth0 |= IB_BTH_SOLICITED;
421 	bth0 |= extra_bytes << 20;
422 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI)
423 		bth0 |= hfi1_get_pkey(ibp, wqe->ud_wr.pkey_index);
424 	else
425 		bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index);
426 	ohdr->bth[0] = cpu_to_be32(bth0);
427 	ohdr->bth[1] = cpu_to_be32(wqe->ud_wr.remote_qpn);
428 	ohdr->bth[2] = cpu_to_be32(mask_psn(wqe->psn));
429 	/*
430 	 * Qkeys with the high order bit set mean use the
431 	 * qkey from the QP context instead of the WR (see 10.2.5).
432 	 */
433 	ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->ud_wr.remote_qkey < 0 ?
434 					 qp->qkey : wqe->ud_wr.remote_qkey);
435 	ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
436 	/* disarm any ahg */
437 	priv->s_ahg->ahgcount = 0;
438 	priv->s_ahg->ahgidx = 0;
439 	priv->s_ahg->tx_flags = 0;
440 	/* pbc */
441 	ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2;
442 
443 	return 1;
444 
445 done_free_tx:
446 	hfi1_put_txreq(ps->s_txreq);
447 	ps->s_txreq = NULL;
448 	return 1;
449 
450 bail:
451 	hfi1_put_txreq(ps->s_txreq);
452 
453 bail_no_tx:
454 	ps->s_txreq = NULL;
455 	qp->s_flags &= ~RVT_S_BUSY;
456 	qp->s_hdrwords = 0;
457 	return 0;
458 }
459 
460 /*
461  * Hardware can't check this so we do it here.
462  *
463  * This is a slightly different algorithm than the standard pkey check.  It
464  * special cases the management keys and allows for 0x7fff and 0xffff to be in
465  * the table at the same time.
466  *
467  * @returns the index found or -1 if not found
468  */
469 int hfi1_lookup_pkey_idx(struct hfi1_ibport *ibp, u16 pkey)
470 {
471 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
472 	unsigned i;
473 
474 	if (pkey == FULL_MGMT_P_KEY || pkey == LIM_MGMT_P_KEY) {
475 		unsigned lim_idx = -1;
476 
477 		for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i) {
478 			/* here we look for an exact match */
479 			if (ppd->pkeys[i] == pkey)
480 				return i;
481 			if (ppd->pkeys[i] == LIM_MGMT_P_KEY)
482 				lim_idx = i;
483 		}
484 
485 		/* did not find 0xffff return 0x7fff idx if found */
486 		if (pkey == FULL_MGMT_P_KEY)
487 			return lim_idx;
488 
489 		/* no match...  */
490 		return -1;
491 	}
492 
493 	pkey &= 0x7fff; /* remove limited/full membership bit */
494 
495 	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i)
496 		if ((ppd->pkeys[i] & 0x7fff) == pkey)
497 			return i;
498 
499 	/*
500 	 * Should not get here, this means hardware failed to validate pkeys.
501 	 */
502 	return -1;
503 }
504 
505 void return_cnp(struct hfi1_ibport *ibp, struct rvt_qp *qp, u32 remote_qpn,
506 		u32 pkey, u32 slid, u32 dlid, u8 sc5,
507 		const struct ib_grh *old_grh)
508 {
509 	u64 pbc, pbc_flags = 0;
510 	u32 bth0, plen, vl, hwords = 5;
511 	u16 lrh0;
512 	u8 sl = ibp->sc_to_sl[sc5];
513 	struct ib_header hdr;
514 	struct ib_other_headers *ohdr;
515 	struct pio_buf *pbuf;
516 	struct send_context *ctxt = qp_to_send_context(qp, sc5);
517 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
518 
519 	if (old_grh) {
520 		struct ib_grh *grh = &hdr.u.l.grh;
521 
522 		grh->version_tclass_flow = old_grh->version_tclass_flow;
523 		grh->paylen = cpu_to_be16((hwords - 2 + SIZE_OF_CRC) << 2);
524 		grh->hop_limit = 0xff;
525 		grh->sgid = old_grh->dgid;
526 		grh->dgid = old_grh->sgid;
527 		ohdr = &hdr.u.l.oth;
528 		lrh0 = HFI1_LRH_GRH;
529 		hwords += sizeof(struct ib_grh) / sizeof(u32);
530 	} else {
531 		ohdr = &hdr.u.oth;
532 		lrh0 = HFI1_LRH_BTH;
533 	}
534 
535 	lrh0 |= (sc5 & 0xf) << 12 | sl << 4;
536 
537 	bth0 = pkey | (IB_OPCODE_CNP << 24);
538 	ohdr->bth[0] = cpu_to_be32(bth0);
539 
540 	ohdr->bth[1] = cpu_to_be32(remote_qpn | (1 << HFI1_BECN_SHIFT));
541 	ohdr->bth[2] = 0; /* PSN 0 */
542 
543 	hdr.lrh[0] = cpu_to_be16(lrh0);
544 	hdr.lrh[1] = cpu_to_be16(dlid);
545 	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
546 	hdr.lrh[3] = cpu_to_be16(slid);
547 
548 	plen = 2 /* PBC */ + hwords;
549 	pbc_flags |= (!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT;
550 	vl = sc_to_vlt(ppd->dd, sc5);
551 	pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
552 	if (ctxt) {
553 		pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL);
554 		if (pbuf)
555 			ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
556 						 &hdr, hwords);
557 	}
558 }
559 
560 /*
561  * opa_smp_check() - Do the regular pkey checking, and the additional
562  * checks for SMPs specified in OPAv1 rev 1.0, 9/19/2016 update, section
563  * 9.10.25 ("SMA Packet Checks").
564  *
565  * Note that:
566  *   - Checks are done using the pkey directly from the packet's BTH,
567  *     and specifically _not_ the pkey that we attach to the completion,
568  *     which may be different.
569  *   - These checks are specifically for "non-local" SMPs (i.e., SMPs
570  *     which originated on another node). SMPs which are sent from, and
571  *     destined to this node are checked in opa_local_smp_check().
572  *
573  * At the point where opa_smp_check() is called, we know:
574  *   - destination QP is QP0
575  *
576  * opa_smp_check() returns 0 if all checks succeed, 1 otherwise.
577  */
578 static int opa_smp_check(struct hfi1_ibport *ibp, u16 pkey, u8 sc5,
579 			 struct rvt_qp *qp, u16 slid, struct opa_smp *smp)
580 {
581 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
582 
583 	/*
584 	 * I don't think it's possible for us to get here with sc != 0xf,
585 	 * but check it to be certain.
586 	 */
587 	if (sc5 != 0xf)
588 		return 1;
589 
590 	if (rcv_pkey_check(ppd, pkey, sc5, slid))
591 		return 1;
592 
593 	/*
594 	 * At this point we know (and so don't need to check again) that
595 	 * the pkey is either LIM_MGMT_P_KEY, or FULL_MGMT_P_KEY
596 	 * (see ingress_pkey_check).
597 	 */
598 	if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE &&
599 	    smp->mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED) {
600 		ingress_pkey_table_fail(ppd, pkey, slid);
601 		return 1;
602 	}
603 
604 	/*
605 	 * SMPs fall into one of four (disjoint) categories:
606 	 * SMA request, SMA response, SMA trap, or SMA trap repress.
607 	 * Our response depends, in part, on which type of SMP we're
608 	 * processing.
609 	 *
610 	 * If this is an SMA response, skip the check here.
611 	 *
612 	 * If this is an SMA request or SMA trap repress:
613 	 *   - pkey != FULL_MGMT_P_KEY =>
614 	 *       increment port recv constraint errors, drop MAD
615 	 *
616 	 * Otherwise:
617 	 *    - accept if the port is running an SM
618 	 *    - drop MAD if it's an SMA trap
619 	 *    - pkey == FULL_MGMT_P_KEY =>
620 	 *        reply with unsupported method
621 	 *    - pkey != FULL_MGMT_P_KEY =>
622 	 *	  increment port recv constraint errors, drop MAD
623 	 */
624 	switch (smp->method) {
625 	case IB_MGMT_METHOD_GET_RESP:
626 	case IB_MGMT_METHOD_REPORT_RESP:
627 		break;
628 	case IB_MGMT_METHOD_GET:
629 	case IB_MGMT_METHOD_SET:
630 	case IB_MGMT_METHOD_REPORT:
631 	case IB_MGMT_METHOD_TRAP_REPRESS:
632 		if (pkey != FULL_MGMT_P_KEY) {
633 			ingress_pkey_table_fail(ppd, pkey, slid);
634 			return 1;
635 		}
636 		break;
637 	default:
638 		if (ibp->rvp.port_cap_flags & IB_PORT_SM)
639 			return 0;
640 		if (smp->method == IB_MGMT_METHOD_TRAP)
641 			return 1;
642 		if (pkey == FULL_MGMT_P_KEY) {
643 			smp->status |= IB_SMP_UNSUP_METHOD;
644 			return 0;
645 		}
646 		ingress_pkey_table_fail(ppd, pkey, slid);
647 		return 1;
648 	}
649 	return 0;
650 }
651 
652 /**
653  * hfi1_ud_rcv - receive an incoming UD packet
654  * @ibp: the port the packet came in on
655  * @hdr: the packet header
656  * @rcv_flags: flags relevant to rcv processing
657  * @data: the packet data
658  * @tlen: the packet length
659  * @qp: the QP the packet came on
660  *
661  * This is called from qp_rcv() to process an incoming UD packet
662  * for the given QP.
663  * Called at interrupt level.
664  */
665 void hfi1_ud_rcv(struct hfi1_packet *packet)
666 {
667 	struct ib_other_headers *ohdr = packet->ohdr;
668 	int opcode;
669 	u32 hdrsize = packet->hlen;
670 	struct ib_wc wc;
671 	u32 qkey;
672 	u32 src_qp;
673 	u16 dlid, pkey;
674 	int mgmt_pkey_idx = -1;
675 	struct hfi1_ibport *ibp = &packet->rcd->ppd->ibport_data;
676 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
677 	struct ib_header *hdr = packet->hdr;
678 	u32 rcv_flags = packet->rcv_flags;
679 	void *data = packet->ebuf;
680 	u32 tlen = packet->tlen;
681 	struct rvt_qp *qp = packet->qp;
682 	bool has_grh = rcv_flags & HFI1_HAS_GRH;
683 	u8 sc5 = hdr2sc(hdr, packet->rhf);
684 	u32 bth1;
685 	u8 sl_from_sc, sl;
686 	u16 slid;
687 	u8 extra_bytes;
688 
689 	qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
690 	src_qp = be32_to_cpu(ohdr->u.ud.deth[1]) & RVT_QPN_MASK;
691 	dlid = be16_to_cpu(hdr->lrh[1]);
692 	bth1 = be32_to_cpu(ohdr->bth[1]);
693 	slid = be16_to_cpu(hdr->lrh[3]);
694 	pkey = (u16)be32_to_cpu(ohdr->bth[0]);
695 	sl = (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xf;
696 	extra_bytes = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
697 	extra_bytes += (SIZE_OF_CRC << 2);
698 	sl_from_sc = ibp->sc_to_sl[sc5];
699 
700 	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
701 	opcode &= 0xff;
702 
703 	process_ecn(qp, packet, (opcode != IB_OPCODE_CNP));
704 	/*
705 	 * Get the number of bytes the message was padded by
706 	 * and drop incomplete packets.
707 	 */
708 	if (unlikely(tlen < (hdrsize + extra_bytes)))
709 		goto drop;
710 
711 	tlen -= hdrsize + extra_bytes;
712 
713 	/*
714 	 * Check that the permissive LID is only used on QP0
715 	 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
716 	 */
717 	if (qp->ibqp.qp_num) {
718 		if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
719 			     hdr->lrh[3] == IB_LID_PERMISSIVE))
720 			goto drop;
721 		if (qp->ibqp.qp_num > 1) {
722 			if (unlikely(rcv_pkey_check(ppd, pkey, sc5, slid))) {
723 				/*
724 				 * Traps will not be sent for packets dropped
725 				 * by the HW. This is fine, as sending trap
726 				 * for invalid pkeys is optional according to
727 				 * IB spec (release 1.3, section 10.9.4)
728 				 */
729 				hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
730 					       pkey, sl,
731 					       src_qp, qp->ibqp.qp_num,
732 					       slid, dlid);
733 				return;
734 			}
735 		} else {
736 			/* GSI packet */
737 			mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey);
738 			if (mgmt_pkey_idx < 0)
739 				goto drop;
740 		}
741 		if (unlikely(qkey != qp->qkey)) {
742 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_Q_KEY, qkey, sl,
743 				       src_qp, qp->ibqp.qp_num,
744 				       slid, dlid);
745 			return;
746 		}
747 		/* Drop invalid MAD packets (see 13.5.3.1). */
748 		if (unlikely(qp->ibqp.qp_num == 1 &&
749 			     (tlen > 2048 || (sc5 == 0xF))))
750 			goto drop;
751 	} else {
752 		/* Received on QP0, and so by definition, this is an SMP */
753 		struct opa_smp *smp = (struct opa_smp *)data;
754 
755 		if (opa_smp_check(ibp, pkey, sc5, qp, slid, smp))
756 			goto drop;
757 
758 		if (tlen > 2048)
759 			goto drop;
760 		if ((hdr->lrh[1] == IB_LID_PERMISSIVE ||
761 		     hdr->lrh[3] == IB_LID_PERMISSIVE) &&
762 		    smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
763 			goto drop;
764 
765 		/* look up SMI pkey */
766 		mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey);
767 		if (mgmt_pkey_idx < 0)
768 			goto drop;
769 	}
770 
771 	if (qp->ibqp.qp_num > 1 &&
772 	    opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
773 		wc.ex.imm_data = ohdr->u.ud.imm_data;
774 		wc.wc_flags = IB_WC_WITH_IMM;
775 		tlen -= sizeof(u32);
776 	} else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
777 		wc.ex.imm_data = 0;
778 		wc.wc_flags = 0;
779 	} else {
780 		goto drop;
781 	}
782 
783 	/*
784 	 * A GRH is expected to precede the data even if not
785 	 * present on the wire.
786 	 */
787 	wc.byte_len = tlen + sizeof(struct ib_grh);
788 
789 	/*
790 	 * Get the next work request entry to find where to put the data.
791 	 */
792 	if (qp->r_flags & RVT_R_REUSE_SGE) {
793 		qp->r_flags &= ~RVT_R_REUSE_SGE;
794 	} else {
795 		int ret;
796 
797 		ret = hfi1_rvt_get_rwqe(qp, 0);
798 		if (ret < 0) {
799 			hfi1_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
800 			return;
801 		}
802 		if (!ret) {
803 			if (qp->ibqp.qp_num == 0)
804 				ibp->rvp.n_vl15_dropped++;
805 			return;
806 		}
807 	}
808 	/* Silently drop packets which are too big. */
809 	if (unlikely(wc.byte_len > qp->r_len)) {
810 		qp->r_flags |= RVT_R_REUSE_SGE;
811 		goto drop;
812 	}
813 	if (has_grh) {
814 		hfi1_copy_sge(&qp->r_sge, &hdr->u.l.grh,
815 			      sizeof(struct ib_grh), 1, 0);
816 		wc.wc_flags |= IB_WC_GRH;
817 	} else {
818 		hfi1_skip_sge(&qp->r_sge, sizeof(struct ib_grh), 1);
819 	}
820 	hfi1_copy_sge(&qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh),
821 		      1, 0);
822 	rvt_put_ss(&qp->r_sge);
823 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
824 		return;
825 	wc.wr_id = qp->r_wr_id;
826 	wc.status = IB_WC_SUCCESS;
827 	wc.opcode = IB_WC_RECV;
828 	wc.vendor_err = 0;
829 	wc.qp = &qp->ibqp;
830 	wc.src_qp = src_qp;
831 
832 	if (qp->ibqp.qp_type == IB_QPT_GSI ||
833 	    qp->ibqp.qp_type == IB_QPT_SMI) {
834 		if (mgmt_pkey_idx < 0) {
835 			if (net_ratelimit()) {
836 				struct hfi1_devdata *dd = ppd->dd;
837 
838 				dd_dev_err(dd, "QP type %d mgmt_pkey_idx < 0 and packet not dropped???\n",
839 					   qp->ibqp.qp_type);
840 				mgmt_pkey_idx = 0;
841 			}
842 		}
843 		wc.pkey_index = (unsigned)mgmt_pkey_idx;
844 	} else {
845 		wc.pkey_index = 0;
846 	}
847 
848 	wc.slid = slid;
849 	wc.sl = sl_from_sc;
850 
851 	/*
852 	 * Save the LMC lower bits if the destination LID is a unicast LID.
853 	 */
854 	wc.dlid_path_bits = dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE) ? 0 :
855 		dlid & ((1 << ppd_from_ibp(ibp)->lmc) - 1);
856 	wc.port_num = qp->port_num;
857 	/* Signal completion event if the solicited bit is set. */
858 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
859 		     (ohdr->bth[0] &
860 		      cpu_to_be32(IB_BTH_SOLICITED)) != 0);
861 	return;
862 
863 drop:
864 	ibp->rvp.n_pkt_drops++;
865 }
866